doi
stringlengths 10
10
| chunk-id
stringlengths 1
4
| chunk
stringlengths 1
1.66k
| id
stringlengths 10
10
| title
stringlengths 19
148
| summary
stringlengths 345
1.92k
| source
stringlengths 31
31
| authors
list | categories
list | comment
stringlengths 4
284
⌀ | journal_ref
stringclasses 14
values | primary_category
stringclasses 16
values | published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2205.11503
|
97
|
i"" 0.52 16.4 43.3 62 0.32 17.2 42.6 56
fg 0.60 13.6 36.1 51 0.46 14.3 32.0 46
–– 0.60 4.4 11.4 20 0.27 3.7 8.9 17
ffgg 0.44 16.2 44.4 50 0.34 17.8 43.0 56
() 0.64 11.3 29.2 46 0.50 15.8 34.1 57
"" 0.57 12.7 34.0 52 0.37 14.8 34.5 43
[] 0.58 13.6 35.3 56 0.51 13.1 29.1 44
hhhiii 0.47 16.1 45.3 58 0.40 13.5 31.7 43
Table 11: Zero-shot performances of the off-the-shelf “small” language models on the original version of the YELPdataset.
Amongst all the model architectures, GPT-J-6B had the finest results, both quantitatively and qualitatively. We also note the
performance differences between the positive to negative direction and the negative to positive direction across all the experiments.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
98
|
performance differences between the positive to negative direction and the negative to positive direction across all the experiments.
It appears that the former direction is easier for all the models than the latter direction. Furthermore, as in the case of AMAZON ,
Table 12 illustrates that most models performed slightly better in the the clean version of Y ELPthan in the original version.
18
Positive!Negative Negative!Positive
Model Delimiter-Pair Acc r-sBLEU s-sBLEU PPL Acc r-sBLEU s-sBLEU PPL
GPT-2-Small
(117M)hi 0.34 13.1 37.9 52 0.14 12.9 38.2 50
* "" 0.38 3.2 9.2 30 0.23 3.7 11.3 29
i"" 0.38 3.1 10.4 25 0.16 5.3 14.9 27
fg 0.36 6.6 19.3 28 0.12 8.2 25.7 29
–– 0.43 4.3 12.4 24 0.12 5.3 16.7 27
ffgg 0.37 16.0 45.0 67 0.12 17.2 54.0 59
() 0.42 8.2 22.5 34 0.13 12.1 37.9 37
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
99
|
() 0.42 8.2 22.5 34 0.13 12.1 37.9 37
"" 0.35 4.9 15.4 33 0.21 6.1 18.5 30
[] 0.42 8.3 23.0 41 0.13 11.3 35.9 39
hhhiii 0.50 1.5 4.0 14 0.11 1.9 5.9 15
GPT-2-Medium
(345M)hi 0.57 8.9 23.7 37 0.31 9.4 27.3 39
* "" 0.64 5.8 14.5 31 0.41 5.4 14.7 29
i"" 0.52 5.7 16.7 28 0.30 5.6 15.0 26
fg 0.65 7.3 19.9 37 0.33 10.5 28.6 41
–– 0.66 3.9 9.7 23 0.34 3.0 7.4 20
ffgg 0.63 13.1 33.5 52 0.31 12.7 35.5 48
() 0.64 9.4 25.3 44 0.29 11.8 33.8 44
"" 0.63 5.2 14.2 29 0.42 6.1 15.9 27
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
100
|
"" 0.63 5.2 14.2 29 0.42 6.1 15.9 27
[] 0.64 7.0 18.4 35 0.33 8.2 22.4 33
hhhiii 0.62 1.9 5.0 15 0.24 1.7 4.8 14
GPT-2-Large
(774M)hi 0.63 14.3 36.3 46 0.27 17.9 48.2 44
* "" 0.65 13.5 33.3 47 0.35 12.5 34.6 36
i"" 0.61 13.9 35.9 47 0.32 15.3 42.9 44
fg 0.67 12.0 28.8 40 0.30 12.5 33.8 30
–– 0.65 5.0 13.7 18 0.26 7.3 19.8 20
ffgg 0.75 17.2 39.9 59 0.31 21.3 58.1 62
() 0.69 12.2 29.2 47 0.31 14.6 40.7 46
"" 0.77 11.8 27.3 41 0.37 11.7 29.6 34
[] 0.75 10.3 24.7 40 0.38 12.9 32.9 38
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
101
|
[] 0.75 10.3 24.7 40 0.38 12.9 32.9 38
hhhiii 0.72 3.6 9.1 16 0.31 4.2 10.7 15
GPT-2-XL
(1558M)hi 0.64 17.4 40.1 58 0.35 17.3 41.5 53
* "" 0.69 11.3 28.2 40 0.41 12.6 31.3 33
i"" 0.71 9.7 22.1 36 0.35 11.6 28.6 34
fg 0.73 8.6 21.3 35 0.46 11.4 25.9 35
–– 0.70 6.0 15.4 23 0.39 6.8 17.5 25
ffgg 0.63 17.4 40.9 70 0.38 19.1 46.8 59
() 0.72 10.8 25.0 45 0.39 14.0 31.1 41
"" 0.77 7.6 17.6 31 0.44 9.9 23.1 30
[] 0.75 10.8 24.9 38 0.41 12.0 29.9 43
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
102
|
[] 0.75 10.8 24.9 38 0.41 12.0 29.9 43
hhhiii 0.68 2.2 5.4 14 0.32 2.0 5.1 13
GPT-Neo-1.3B
(1.3B)hi 0.68 6.5 16.7 27 0.42 6.9 17.6 29
* "" 0.38 12.5 37.0 37 0.22 12.5 36.2 33
i"" 0.32 13.7 42.3 41 0.19 16.1 47.9 40
fg 0.69 4.6 10.5 22 0.37 6.3 15.3 23
–– 0.58 3.1 8.1 18 0.33 4.2 11.1 17
ffgg 0.69 7.2 17.0 30 0.40 8.9 20.6 27
() 0.63 8.6 21.3 39 0.28 8.3 23.3 28
"" 0.47 12.6 35.2 43 0.30 13.8 35.6 36
[] 0.68 8.3 21.4 40 0.34 8.6 23.9 30
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
103
|
[] 0.68 8.3 21.4 40 0.34 8.6 23.9 30
hhhiii 0.72 1.2 2.9 15 0.38 1.7 3.6 15
GPT-Neo-2.7B
(2.7B)hi 0.66 8.8 23.4 31 0.32 13.9 35.3 36
* "" 0.58 14.5 36.9 42 0.17 17.4 51.1 42
i"" 0.54 13.8 38.3 43 0.21 13.2 39.9 32
fg 0.64 7.0 19.0 24 0.28 11.0 32.6 29
–– 0.68 3.4 9.1 17 0.26 5.0 14.3 16
ffgg 0.57 11.0 29.3 31 0.24 15.5 46.1 35
() 0.76 10.3 23.1 44 0.41 14.4 33.6 43
"" 0.59 12.5 33.6 42 0.21 14.9 43.4 38
[] 0.66 9.6 23.7 32 0.29 14.9 42.4 43
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
104
|
[] 0.66 9.6 23.7 32 0.29 14.9 42.4 43
hhhiii 0.64 5.4 14.4 21 0.27 8.1 21.7 23
GPT-J-6B
(6B)hi 0.62 14.1 35.3 50 0.47 14.7 33.4 44
* "" 0.55 17.1 43.9 65 0.40 13.2 31.8 41
i"" 0.61 16.9 41.6 56 0.38 13.3 30.5 37
fg 0.61 14.3 34.7 49 0.48 13.5 30.6 43
–– 0.54 5.4 14.7 22 0.36 4.8 10.9 19
ffgg 0.42 14.7 42.3 38 0.33 17.5 45.0 53
() 0.66 12.9 30.7 50 0.51 11.5 23.4 44
"" 0.66 15.7 36.1 55 0.40 16.6 36.2 45
[] 0.69 11.8 28.7 45 0.53 13.3 27.3 43
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
105
|
[] 0.69 11.8 28.7 45 0.53 13.3 27.3 43
hhhiii 0.53 10.7 29.8 35 0.43 11.4 25.4 30
Table 12: Zero-shot performances of the off-the-shelf “small” language models on the clean version of the YELP dataset
(YELP-clean, in short). In contrast to Table 11, we note that models, overall, achieved better results in YELP-clean than in
YELP-original. Some models even could go beyond the 75% accuracy level in the positive to negative direction. Consistent with
the previous findings, these results also indicate that curly brackets fg, square brackets [], parentheses (), and quotes ""are
favourable delimiter-pairs, leading to better outcomes than many other delimiter-pairs.
19
Shakespearean!Modern English
Model Setting Acc r-sBLEU s-sBLEU PPL
GPT-2-Small (117M) 4-Shot (Top Choice) 0.35 17.1 42.4 65
GPT-2-Medium (345M) 4-Shot (Top Choice) 0.50 7.1 13.9 65
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
106
|
GPT-2-Medium (345M) 4-Shot (Top Choice) 0.50 7.1 13.9 65
GPT-2-Large (774M) 4-Shot (Top Choice) 0.38 14.1 30.9 134
GPT-2-XL (1558M) 4-Shot (Top Choice) 0.39 18.9 38.4 90
GPT-Neo-1.3B (1.3B) 4-Shot (Top Choice) 0.39 17.2 37.0 63
GPT-Neo-2.7B (2.7B) 4-Shot (Top Choice) 0.62 23.9 41.4 106
GPT-J-6B (6B) 4-Shot (Top Choice) 0.78 21.9 31.8 81
Table 13: Four-shot performances of the off-the-shelf “small” language models on the clean version of the SHAKESPEARE
corpus. In this few-shot setup, we included a simple natural-language task description and four illustrative examples in the
prompt. We note that GPT-J-6B was able to “translate” sentences written in Elizabethan English to Modern English successfully,
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
107
|
prompt. We note that GPT-J-6B was able to “translate” sentences written in Elizabethan English to Modern English successfully,
achieving a transfer accuracy score of 78% , reference BLEU score of 21.9, and perplexity value of 81.
Positive!Negative Negative!Positive
Model Setting Acc r-sBLEU s-sBLEU PPL Acc r-sBLEU s-sBLEU PPL
Style-Embedding
Li et al. (2018)0.33 16.2 33.2 265 0.47 13.1 29.0 287
CrossAligned 0.66 2.2 3.0 93 0.74 1.7 2.4 96
DeleteAndRetrieve 0.49 33.3 60.3 120 0.51 26.7 53.5 113
TemplateBased 0.65 38.1 70.5 243 0.56 31.0 65.7 200
GPT-2-Small
(117M)0-Shot (Top Choice) 0.31 18.9 34.9 49 0.18 17.7 38.1 48
4-Shot (Top Choice) 0.32 23.8 46.1 94 0.25 27.7 58.2 67
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
108
|
4-Shot (Top Choice) 0.32 23.8 46.1 94 0.25 27.7 58.2 67
4-Shot (RC: 3, IF) 0.42 21.1 39.2 68 0.32 24.6 51.0 69
4-Shot (RC: 3, FS) 0.38 23.0 43.5 73 0.30 27.2 52.4 77
GPT-2-Medium
(345M)0-Shot (Top Choice) 0.48 21.9 36.7 68 0.32 20.1 38.0 57
4-Shot (Top Choice) 0.44 12.3 17.8 78 0.42 11.5 17.8 72
4-Shot (RC: 3, IF) 0.58 12.6 18.6 66 0.55 10.2 15.0 53
4-Shot (RC: 3, FS) 0.52 7.2 10.0 59 0.50 5.6 9.0 56
GPT-2-Large
(774M)0-Shot (Top Choice) 0.44 28.6 49.1 70 0.28 26.0 51.2 55
4-Shot (Top Choice) 0.47 17.1 27.7 54 0.32 15.0 27.4 101
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
109
|
4-Shot (Top Choice) 0.47 17.1 27.7 54 0.32 15.0 27.4 101
4-Shot (RC: 3, IF) 0.60 15.4 24.7 62 0.43 15.6 27.7 59
4-Shot (RC: 3, FS) 0.55 21.6 33.1 55 0.35 20.0 34.7 53
GPT-2-XL
(1558M)0-Shot (Top Choice) 0.46 21.5 35.4 59 0.32 22.3 41.4 70
4-Shot (Top Choice) 0.63 13.7 20.3 65 0.44 14.5 22.3 60
4-Shot (RC: 3, IF) 0.70 11.5 17.2 77 0.56 13.2 19.9 50
4-Shot (RC: 3, FS) 0.66 11.5 16.6 54 0.50 14.8 19.7 58
GPT-Neo-1.3B
(1.3B)0-Shot (Top Choice) 0.49 11.8 19.9 34 0.31 10.9 20.5 35
4-Shot (Top Choice) 0.53 22.1 35.8 68 0.34 22.0 39.6 67
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
110
|
4-Shot (Top Choice) 0.53 22.1 35.8 68 0.34 22.0 39.6 67
4-Shot (RC: 3, IF) 0.60 21.2 33.4 66 0.39 21.6 36.2 65
4-Shot (RC: 3, FS) 0.56 22.2 33.1 67 0.32 20.9 32.6 63
GPT-Neo-2.7B
(2.7B)0-Shot (Top Choice) 0.48 21.4 36.5 55 0.28 23.7 45.9 57
4-Shot (Top Choice) 0.52 22.3 33.7 74 0.35 22.3 39.5 74
4-Shot (RC: 3, IF) 0.60 21.7 32.3 69 0.42 20.6 34.9 66
4-Shot (RC: 3, FS) 0.55 21.2 30.2 65 0.40 19.2 29.7 60
GPT-J-6B
(6B)0-Shot (Top Choice) 0.41 26.1 44.6 57 0.33 27.1 47.7 72
4-Shot (Top Choice) 0.59 20.5 31.9 69 0.46 18.1 28.8 60
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
111
|
4-Shot (Top Choice) 0.59 20.5 31.9 69 0.46 18.1 28.8 60
4-Shot (RC: 3, IF) 0.65 21.5 31.4 70 0.52 19.3 29.3 58
4-Shot (RC: 3, FS) 0.64 17.0 24.7 61 0.50 18.6 25.7 59
Table 14: Four-shot results on AMAZON -clean. We show the average results under (i) the zero-shot setting ( 0-Shot (Top
Choice) ); (ii) the four-shot setting in which we chose the top beam search result ( 4-Shot (Top Choice) ; (iii) the four-shot setting
in which we generated three outputs from the model, re-scored and re-ranked them according to the textual similarity and style
factors—ignoring the fluency aspect—( 4-Shot (RC: 3, IF) ); and (iv) the four-shot setting in which we generated three outputs
from the model, re-scored and re-ranked them according to the textual similarity, style, and fluency factors ( 4-Shot (RC: 3, FS) .
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
112
|
from the model, re-scored and re-ranked them according to the textual similarity, style, and fluency factors ( 4-Shot (RC: 3, FS) .
(Here, “RC” denotes to the re-ranking-and-choosing method, “IF” ignoring fluency, and “FS” full set (meaning that we consider
all the textual similarity, transfer accuracy, and fluency criteria). First, we stress that proving few-shot examples in the input
resulted in 10-15% improvements in the accuracy scores in most of our models (see 0-shot results vs. 4-shot results). Second, we
highlight that some of our off-the-shelf models (e.g., GPT-2-XL and GPT-J-6B) performed on par with, and even succeeded
the performances of, the specially-tailored models of Li et al. (2018) along certain metrics. (For instance, our off-the-shelf
models achieve significantly lower perplexity rates than theirs.) Third, we note that the Prompt-and-Rerank method (described in
§3.2) seems to boost the models’ performances in almost all the cases. Fourth, we note that 4-Shot (RC: 3, IF) often performs
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
113
|
§3.2) seems to boost the models’ performances in almost all the cases. Fourth, we note that 4-Shot (RC: 3, IF) often performs
noticeably better than 4-Shot (RC: 3, FS) across all the models, suggesting that we may not need to include the fluency factor in
our re-scoring calculations after all.
20
Positive!Negative Negative!Positive
Model Setting Acc r-sBLEU s-sBLEU PPL Acc r-sBLEU s-sBLEU PPL
BackTranslation (Prabhumoye et al., 2018) 0.90 2.0 2.7 120 0.99 1.9 2.6 64
UnpairedRL (Xu et al., 2018) 0.42 16.1 46.0 408 0.56 17.5 45.3 362
CrossAlignment (Shen et al., 2017) 0.72 7.3 19.3 244 0.74 8.3 19.3 190
Multidecoder(Fu et al., 2018)0.42 13.4 43.2 376 0.49 12.6 35.5 369
StyleEmbedding 0.08 19.7 71.3 154 0.10 18.9 62.7 197
Style-Embedding
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
114
|
StyleEmbedding 0.08 19.7 71.3 154 0.10 18.9 62.7 197
Style-Embedding
Li et al. (2018)0.08 19.7 71.3 154 0.10 18.9 62.7 197
Delete-Only 0.89 12.7 33.1 195 0.81 14.0 34.7 169
Retrieve-Only 1.00 1.1 2.1 93 0.98 1.8 2.8 86
CrossAligned 0.72 7.3 19.3 244 0.74 8.3 19.3 190
DeleteAndRetrieve 0.90 14.5 36.8 279 0.89 14.8 35.9 100
TemplateBased 0.84 21.2 55.2 289 0.83 20.9 55.7 190
DualR (Luo et al., 2019b) 0.91 26.5 58.7 125 0.85 25.3 58.8 141
B-GST (Sudhakar et al., 2019) 0.83 19.8 46.8 153 0.79 23.4 46.1 163
Multi-Class (StyleTransformer ,
Dai et al. (2019) )0.94 26.3 61.0 177 0.77 26.5 65.0 173
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
115
|
Multi-Class (StyleTransformer ,
Dai et al. (2019) )0.94 26.3 61.0 177 0.77 26.5 65.0 173
Conditional 0.95 22.6 52.6 211 0.87 23.1 53.0 234
GPT-2-Small
(117M)0-Shot (Top Choice) 0.36 6.6 19.3 28 0.12 8.2 25.7 29
4-Shot (Top Choice) 0.08 24.8 75.2 94 0.10 23.1 74.1 81
4-Shot (RC: 3, IF) 0.14 21.5 71.0 73 0.17 18.5 58.5 99
4-Shot (RC: 3, FS) 0.06 26.7 84.6 72 0.09 27.9 85.4 68
GPT-2-Medium
(345M)0-Shot (Top Choice) 0.65 7.3 19.9 37 0.33 10.5 28.6 41
4-Shot (Top Choice) 0.49 14.5 34.1 72 0.35 13.3 35.2 56
4-Shot (RC: 3, IF) 0.68 15.0 35.1 69 0.53 12.6 29.8 45
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
116
|
4-Shot (RC: 3, IF) 0.68 15.0 35.1 69 0.53 12.6 29.8 45
4-Shot (RC: 3, FS) 0.43 20.4 46.4 74 0.40 17.7 43.5 48
GPT-2-Large
(774M)0-Shot (Top Choice) 0.67 12.0 28.8 40 0.30 12.5 33.8 30
4-Shot (Top Choice) 0.79 16.6 32.8 84 0.57 14.5 31.0 74
4-Shot (RC: 3, IF) 0.79 10.6 24.7 79 0.58 12.1 30.3 53
4-Shot (RC: 3, FS) 0.58 23.0 56.8 76 0.45 22.1 53.5 64
GPT-2-XL
(1558M)0-Shot (Top Choice) 0.73 8.6 21.3 35 0.46 11.4 25.9 35
4-Shot (Top Choice) 0.63 13.7 20.3 65 0.44 14.5 22.3 60
4-Shot (RC: 3, IF) 0.87 14.8 28.7 65 0.72 12.0 25.3 55
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
117
|
4-Shot (RC: 3, IF) 0.87 14.8 28.7 65 0.72 12.0 25.3 55
4-Shot (RC: 3, FS) 0.77 21.1 38.7 85 0.62 18.0 35.2 70
GPT-Neo-1.3B
(1.3B)0-Shot (Top Choice) 0.69 4.6 10.5 22 0.37 6.3 15.3 23
4-Shot (Top Choice) 0.78 14.8 30.2 58 0.45 14.5 32.0 56
4-Shot (RC: 3, IF) 0.85 14.6 30.1 59 0.61 13.1 28.3 42
4-Shot (RC: 3, FS) 0.77 22.5 46.1 87 0.49 23.5 46.1 72
GPT-Neo-2.7B
(2.7B)0-Shot (Top Choice) 0.64 7.0 19.0 24 0.28 11.0 32.6 29
4-Shot (Top Choice) 0.83 22.8 42.7 89 0.42 21.8 47.1 89
4-Shot (RC: 3, IF) 0.88 23.5 45.8 96 0.52 22.0 48.0 69
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
118
|
4-Shot (RC: 3, IF) 0.88 23.5 45.8 96 0.52 22.0 48.0 69
4-Shot (RC: 3, FS) 0.80 24.5 44.5 87 0.48 23.9 48.4 68
GPT-J-6B
(6B)0-Shot (Top Choice) 0.61 14.3 34.7 49 0.48 13.5 30.6 43
4-Shot (Top Choice) 0.81 25.3 50.5 107 0.52 21.7 48.7 82
4-Shot (RC: 3, IF) 0.87 23.0 47.7 80 0.65 20.2 44.6 58
4-Shot (RC: 3, FS) 0.79 25.9 51.5 78 0.55 26.3 50.0 67
Table 15: Four-shot results on YELP-clean. As before, we detail the average results under different zero- and few-shot settings:
Overall, our few-shot results on YELP-clean are consistent with those on AMAZON -clean, as reported in Table 14. GPT-2-XL
and GPT-J-6B models, amongst all the models, have achieved the most successful performances, leveling themselves almost with
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
119
|
and GPT-J-6B models, amongst all the models, have achieved the most successful performances, leveling themselves almost with
the custom-made (trained) state-of-the-art models. We present some of the generated examples from these models in Table 20.
Ungrammatical!Grammatical
Model Setup GLEU r-sBLEU s-sBLEU PPL
GPT-2-Small (117M) 4-Shot (Top Choice) 35.9 74.8 91.5 76
GPT-2-Medium (345M) 4-Shot (Top Choice) 19.9 38.0 40.6 63
GPT-2-Large (774M) 4-Shot (Top Choice) 30.0 56.8 64.1 55
GPT-2-XL (1558M) 4-Shot (Top Choice) 24.8 46.2 47.0 57
GPT-Neo-1.3B (1.3B) 4-Shot (Top Choice) 26.6 48.4 49.4 54
GPT-Neo-2.7B (2.7B) 4-Shot (Top Choice) 34.5 57.4 54.1 40
GPT-J-6B (6B) 4-Shot (Top Choice) 40.0 64.8 59.1 48
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
120
|
GPT-J-6B (6B) 4-Shot (Top Choice) 40.0 64.8 59.1 48
Table 16: Four-shot performances of the off-the-shelf “small” language models on the clean version of the JFLEG corpus. In
this task, as a baseline, we consider the model which directly copies its input—we call this model “copy-input” model; this model
achieves a GLEU score score 37.7. All but GPT-J-6B fail to beat the performance of the baseline “copy-input” model. GPT-J-6B,
on the other hand, achieves a GLEU score of 40.0. Small language models fail, in fact rather miserably, at this grammatical error
correction task. There is therefore an open room for improvement. We hope that our results will encourage researchers to come
up with more effective ways to utilize pre-trained language models to solve this challenging problem.
21
Informal!Formal
Model Setup Accuracy r-sBLEU s-sBLEU PPL
GPT-2-Small (117M) 4-Shot (Top Choice) 0.85 6.1 8.7 41
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
121
|
GPT-2-Small (117M) 4-Shot (Top Choice) 0.85 6.1 8.7 41
GPT-2-Medium (345M) 4-Shot (Top Choice) 0.76 12.9 16.2 39
GPT-2-Large (774M) 4-Shot (Top Choice) 0.78 23.2 31.3 33
GPT-2-XL (1558M) 4-Shot (Top Choice) 0.82 32.7 41.9 58
GPT-Neo-1.3B (1.3B) 4-Shot (Top Choice) 0.85 36.4 49.6 68
GPT-Neo-2.7B (2.7B) 4-Shot (Top Choice) 0.81 50.0 61.2 64
GPT-J-6B (6B) 4-Shot (Top Choice) 0.69 47.9 52.3 49
Table 17: Four-shot results on GYAFC -clean. We highlight that most of the off-the-shelf “small” language models could
obtain at least 80% accuracy in the informal to formal direction. Amongst all the models, GPT-2-XL, GPT-Neo-1.3B, and
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
122
|
obtain at least 80% accuracy in the informal to formal direction. Amongst all the models, GPT-2-XL, GPT-Neo-1.3B, and
GPT-Neo-2.7B appeared to be most successful, achieving not only high accuracy scores but also high BLEU scores and relatively
low perplexity rates.
Model Correct-Class Accuracy Opposite-Class Accuracy reference -sBLEU
GPT-2-Small 0.42 0.46 51.9
GPT-2-Medium 0.46 0.46 60.3
GPT-2-Large 0.53 0.35 65.6
GPT-2-XL 0.56 0.38 68.5
GPT-Neo-1.3B 0.55 0.37 67.3
GPT-Neo-2.7B 0.57 0.38 69.6
GPT-J-6B 0.74 0.21 81.9
Table 18: Four-shot performances of the off-the-shelf “small” language models on the symbolic manipulation task ( SYM)
defined in §5.1. Correct-Class accuracy refers to the accuracy of the model under exact-string matching, whereas Opposite-Class
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
123
|
defined in §5.1. Correct-Class accuracy refers to the accuracy of the model under exact-string matching, whereas Opposite-Class
accuracy refers to the fraction of the cases for which the model copied and placed the right input words in the output but
verbalized the incorrect (opposite) inequality symbol, that is writing “less than” instead of “greater than” or vice versa in between
the expressions (for instance, the ground-truth might be “olive is greater than cat”, but the model might have generated “olive is
less than cat.”) It was surprising to discover that most models failed to go beyond 60% accuracy on this small dataset. GPT-J-6B,
on the other hand, outperformed all the other models, achieving an accuracy score of 74% on this task. We also remark that of
the cases for which the models failed to generate the correct output, they often were able to copy the appropriate words from the
input but failed to write the correct inequality symbol at the end.
Positive!Negative Negative!Positive
Model Setting Acc r-sBLEU s-sBLEU PPL Acc r-sBLEU s-sBLEU PPL
GPT-2-Small
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
124
|
Model Setting Acc r-sBLEU s-sBLEU PPL Acc r-sBLEU s-sBLEU PPL
GPT-2-Small
(117M)4-Shot, Vanilla 0.14 25.3 82.1 79 0.13 24.9 80.6 75
4-Shot, Contrastive 0.14 21.5 71. 0 73 0.17 18.5 58.5 99
4-Shot, Negation-v1 0.05 25.2 83.7 72 0.07 23.1 75.3 68
4-Shot, Negation-v2 0.06 25.4 84.5 75 0.08 25.3 83.0 74
GPT-2-Medium
(345M)4-Shot, Vanilla 0.63 19.7 49.4 75 0.38 17.7 49.7 62
4-Shot, Contrastive 0.68 15.0 35.1 69 0.53 12.6 29.8 45
4-Shot, Negation-v1 0.34 15.7 41.6 63 0.22 15.4 42.3 55
4-Shot, Negation-v2 0.36 14.1 39.2 63 0.30 12.6 36.2 51
GPT-2-Large
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
125
|
4-Shot, Negation-v2 0.36 14.1 39.2 63 0.30 12.6 36.2 51
GPT-2-Large
(774M)4-Shot, Vanilla 0.75 16.2 38.9 60 0.52 17.2 45.6 55
4-Shot, Contrastive 0.79 10.6 24.7 79 0.58 12.1 30.3 53
4-Shot, Negation-v1 0.41 10.8 30.1 79 0.27 13.4 36.7 59
4-Shot, Negation-v2 0.14 16.9 52.1 57 0.22 12.6 36.3 59
GPT-2-XL
(1558M)4-Shot, Vanilla 0.86 15.6 32.0 59 0.70 13.8 29.9 58
4-Shot, Contrastive 0.87 14.8 28.7 65 0.72 12.0 25.3 55
4-Shot, Negation-v1 0.83 11.8 24.1 81 0.50 14.9 32.0 54
4-Shot, Negation-v2 0.53 19.0 43.5 77 0.51 16.9 37.7 61
GPT-Neo-1.3B
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
126
|
4-Shot, Negation-v2 0.53 19.0 43.5 77 0.51 16.9 37.7 61
GPT-Neo-1.3B
(1.3B)4-Shot, Vanilla 0.80 17.2 38.5 80 0.52 14.5 35.6 50
4-Shot, Contrastive 0.85 14.6 30.1 59 0.61 13.1 28.3 42
4-Shot, Negation-v1 0.79 16.1 34.7 72 0.00 0.0 0.0 0
4-Shot, Negation-v2 0.57 16.5 40.6 67 0.00 0.0 0.0 0
GPT-Neo-2.7B
(2.7B)4-Shot, Vanilla 0.86 24.7 51.2 104 0.43 24.3 54.9 74
4-Shot, Contrastive 0.88 23.5 45.8 96 0.52 22.0 48.0 69
4-Shot, Negation-v1 0.80 22.5 47.8 79 0.00 0.0 0.0 0
4-Shot, Negation-v2 0.76 22.0 48.2 85 0.00 0.0 0.0 0
GPT-J-6B
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
127
|
4-Shot, Negation-v2 0.76 22.0 48.2 85 0.00 0.0 0.0 0
GPT-J-6B
(6B)4-Shot, Vanilla 0.90 23.5 51.0 85 0.62 22.8 49.9 63
4-Shot, Contrastive 0.87 23.0 47.7 80 0.65 20.2 44.6 58
4-Shot, Negation-v1 0.82 23.6 50.9 85 0.53 24.3 54.7 65
4-Shot, Negation-v2 0.73 25.1 59.1 95 0.48 25.2 58.1 66
Table 19: Four-shot performances of the off-the-shelf “small” language models on the YELP-clean dataset under different
prompting protocols. Rows are bolded according to best accuracy scores. Across all models, the vanilla and contrastive prompting
protocols appear to yield the most favourable results, though the contrastive prompting protocol typically performs better than
the vanilla prompting protocol along the accuracy dimension, but not necessarily along the r-sBLEU dimension. We, however,
note that in the case of GPT-J-6B, vanilla prompting performs on par with contrastive prompting overall.
22
Model Text
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
128
|
note that in the case of GPT-J-6B, vanilla prompting performs on par with contrastive prompting overall.
22
Model Text
Input (P!N) the staff is awesome and location is right in the heart of old town!
GPT-2-Small the staff is awesome and location is right in the heart of old town!
GPT-2-Medium the staff was great but the location was an even higher priority.
GPT-2-Large the staff is awful and location is not in the center of old town!
GPT-2-XL the staff was kind of awful, location was kind of a shithole
GPT-Neo-1.3B the staff is awful and location is just a couple minutes away from the city center!
GPT-Neo-2.7B the staff is terrible and location is just outside of old town!
GPT-J-6B the staff is terrible and location is in the middle of nowhere!
Input (P!N) he always looks gorgeous and is so happy when i pick him up!
GPT-2-Small he always looks gorgeous and is so happy when i pick him up!
GPT-2-Medium he always looks so sad and upset by the fact that i pick him up!
GPT-2-Large he always looks ugly and is so sad when i pick him up!
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
129
|
GPT-2-Large he always looks ugly and is so sad when i pick him up!
GPT-2-XL he never looks gorgeous and is so sad when i pick him up!
GPT-Neo-1.3B he always looks awful and is so sad when i pick him up!
GPT-Neo-2.7B he always looks horrible and is so unhappy when i pick him up!
GPT-J-6B he always looks ugly and is so sad when i pick him up!
Input (N!P) the food was low quality and thrown together quickly with little flavor.
GPT-2-Small the food was low quality and thrown together quickly with little flavor.
GPT-2-Medium the food was high quality and thrown together quickly with a great flavor.
GPT-2-Large the food was excellent and was cooked to perfection by very skilled and professional server-chefs.
GPT-2-XL the food was high quality and tasted really good.
GPT-Neo-1.3B the food was high quality and tasted amazing!
GPT-Neo-2.7B the food was fantastic and presented exceptionally with excellent quality and excellent flavors!
GPT-J-6B the food was delicious and prepared with a lot of care!
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
130
|
GPT-J-6B the food was delicious and prepared with a lot of care!
Input (N!P) we went there on a busy holiday weekend and the service was extremely slow.
GPT-2-Small we went there on a busy holiday weekend and the service was extremely slow.
GPT-2-Medium we went there on a busy holiday weekend and the service was exceptional!
GPT-2-Large we did go there on a busy holiday weekend–it was fast!
GPT-2-XL we went there on a quiet weekend and the service was very fast.
GPT-Neo-1.3B we went there on a busy holiday weekend and the service was exceptional!
GPT-Neo-2.7B we went there on a slow holiday weekend and the service was extremely fast
GPT-J-6B we went there on a busy holiday weekend and the service was extremely fast.
Input (P!N) i love the outside seating and the reverse happy hour is incredibly affordable.
GPT-2-Small i love the outside seating and the reverse happy hour is unbelievably affordable.
GPT-2-Mediumi love everything about this place–the music is great, everyone is super nice, and they have
amazing desserts that are affordable.
GPT-2-Large i loved the outside seating and the reverse happy hour is insanely expensive.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
131
|
amazing desserts that are affordable.
GPT-2-Large i loved the outside seating and the reverse happy hour is insanely expensive.
GPT-2-XL i hate the outside seating and the reverse happy hour.
GPT-Neo-1.3B i hate the outside seating and the reverse happy hour is incredibly costly.
GPT-Neo-2.7B i hate the outside seating and the reverse happy hour is outrageously expensive.
GPT-J-6B i hate the outside seating and the reverse happy hour is incredibly expensive.
Input (N!P) the service is mediocre and the restaurant is just average.
GPT-2-Small i love their black tea and hot chocolate selections!
GPT-2-Mediumi liked their black tea and hot chocolate selections; i was impressed by the quality of the food
and service
GPT-2-Large the service is wonderful!
GPT-2-XL the service is great, and the restaurant is great!
GPT-Neo-1.3B the service and the restaurant are just average.
GPT-Neo-2.7B the service and the restaurant are just wonderful!
GPT-J-6B the service is remarkable and the restaurant is amazing.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
132
|
GPT-Neo-2.7B the service and the restaurant are just wonderful!
GPT-J-6B the service is remarkable and the restaurant is amazing.
Table 20: Qualitative examples of sentiment transfer on YELP-clean. We manually highlight in blue and red phrases that we find
to be associated with the input and target styles, respectively.
23
Model Text
Input (N!P) it works for maybe a few minutes and wears off quickly.
GPT-2-Small it does not wear or work well when driving.
GPT-2-Medium i think that it may have something to do with this brand.
GPT-2-Large it works well for a while, but it wears off very quickly.
GPT-2-XL it only does a few minutes’ worth of work and then gives up.
GPT-Neo-1.3B it can be used for a few minutes, but wears off quickly.
GPT-Neo-2.7B it lasts a lifetime and wears off after a few hours.
GPT-J-6B it works for a few hours and wears well.
Input (N!P) i am only giving this item num_num stars because of the absurd price.
GPT-2-Small i am only giving this item num_num stars because of the absurd price.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
133
|
GPT-2-Small i am only giving this item num_num stars because of the absurd price.
GPT-2-Medium i will buy some num_num stars and will use them to have an extra day off when i work.
GPT-2-Large i am only giving this item stars because it got a good review.
GPT-2-XL i will NOT purchase this item again.
GPT-Neo-1.3B i really liked it and i will do it more often.
GPT-Neo-2.7B i am giving this item num_num stars, because of the absurd price.
GPT-J-6B i am only giving this item 5 stars because it works perfect and the price is reasonable.
Input (N!P) would have been alot nicer if it simply threaded in with the correct alignment.
GPT-2-Small would have been alot nicer if it simply threaded in with the correct alignment.
GPT-2-Medium it works perfectly.
GPT-2-Large would have been so much better if the plastic was a bit thicker.
GPT-2-XL would have been great if it just threaded in properly.
GPT-Neo-1.3B would have been alot nicer if it simply threaded in with the correct alignment.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
134
|
GPT-Neo-1.3B would have been alot nicer if it simply threaded in with the correct alignment.
GPT-Neo-2.7B it is very nice but it does not fit in properly at all.
GPT-J-6B it works just fine and there is zero frustration with the way the threading is implemented.
Input (P!N) it fulfills my modest requirements, and exceeds my expectations in appearance.
GPT-2-Small it isn’t good enough. I would buy from this brand again.
GPT-2-Medium it doesn’t fit my needs and falls short in other areas.
GPT-2-Largeit did not fulfill my modest requirements, and did not exceed my expectations in appearance, and
would not fit well in the car.
GPT-2-XL it does not meet my expectations in appearance.
GPT-Neo-1.3B it does not fit my expectations and it does not fulfill my modest requirements in appearance.
GPT-Neo-2.7B it doesn t fulfill my modest requirements, and i found it to be unsatisfactory.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
135
|
GPT-Neo-2.7B it doesn t fulfill my modest requirements, and i found it to be unsatisfactory.
GPT-J-6B it fails to meet my needs, and is disappointing to look at.
Input (P!N) great headset, solid build quality and blocks out the wind.
GPT-2-Small i hate it
GPT-2-Medium lots of dust on here and can be worn.
GPT-2-Large not worth buying and not worth the price. It’s just too large.
GPT-2-XL i hate it, it is terrible.
GPT-Neo-1.3B dreadful headset, terrible quality and wind blocks out the audio quality.
GPT-Neo-2.7B lousy headset, hard to unblock the wind.
GPT-J-6B great headset, but it does not work great in the wind.
Input (P!N) i dropped phone once and the case held up perfectly.
GPT-2-Small i dropped phone once and the case held up perfectly.
GPT-2-Medium i dropped phone as a result.
GPT-2-Large i never dropped my phone once and the case held up fine.
GPT-2-XL i dropped phone twice and the case was damaged.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
136
|
GPT-2-Large i never dropped my phone once and the case held up fine.
GPT-2-XL i dropped phone twice and the case was damaged.
GPT-Neo-1.3B i dropped it and it was very broken.
GPT-Neo-2.7B i dropped phone as many times as I dropped a brick.
GPT-J-6B i dropped the phone and the case broke into pieces.
Table 21: Qualitative examples of sentiment transfer on AMAZON -clean. We manually highlight in blue and red phrases that we
find to be associated with the input and target styles, respectively.
24
Model Text
Input follow ur heart... it’s neva wrong
GPT-2-Small follow my heart.
GPT-2-Large follow your heart... its nova wrong...
GPT-2-Medium follow your heart... its nova wrong
GPT-2-XL follow ur heart and be the man you always wanted to be and never had.
GPT-J-6B follow your heart... it is never wrong
GPT-Neo-1.3B follow your heart... its neva wrong
GPT-Neo-2.7B follow your heart... it’s never wrong
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
137
|
GPT-Neo-1.3B follow your heart... its neva wrong
GPT-Neo-2.7B follow your heart... it’s never wrong
Input if u you act like dat all da time she might think u desperate.
GPT-2-Smallif u acts like dat all, she might think that u wanna be happy with something and u dont want dat
to break up with you, you will regret it, and she will become jealous. it is good...
GPT-2-Medium if u u are always like dat..
GPT-2-Large if you act like dat all da time she might think you desperate...
GPT-2-XL if u u act like dat all da time she might think u desperate :)
GPT-Neo-1.3B if u you act like dat everytime she thinks u desperately searching.
GPT-Neo-2.7B if u you act like dat all da time she might think u desperate...
GPT-J-6B if you act like that all the time she might think you desperate
Input dont do things tht can get you hurt.
GPT-2-Small but im sorry that u can do it and u can’t leave
GPT-2-Medium do things which i’m sure u will like.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
138
|
GPT-2-Small but im sorry that u can do it and u can’t leave
GPT-2-Medium do things which i’m sure u will like.
GPT-2-Large don’t do things that can get you hurt
GPT-2-XL dont do things that can end up getting u hurt.
GPT-Neo-1.3B do things that can get you hurt.
GPT-Neo-2.7B do not do things that can get you hurt.
GPT-J-6B dont do things that might hurt you
Table 22: Qualitative examples of sentiment transfer on GYAFC -clean. We manually highlight in blue and red phrases that we
find to be associated with the input and target styles, respectively.
Model Text
InputFo example, in this genaration people watch movies and listen to musics more than read books
and novels.
GPT-2-SmallFo example, in this genaration people watch movies and listen to musics more than read books
and novels.
GPT-2-Medium In this generation people watch movies more than read books and novels.
GPT-2-LargeFo example, in this genaration people watch movies and listen to musics more than read books
and novels.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
139
|
GPT-2-LargeFo example, in this genaration people watch movies and listen to musics more than read books
and novels.
GPT-2-XL Fo example, in this example people watch movies and listen to music more than read books.
GPT-Neo-1.3BExample, in this genaration people watch movies and listen to musics more than read books and
novels.
GPT-Neo-2.7BExamples, in this generation, people watch movies and listen to music, whereas, reading books
and novels is limited.
GPT-J-6BFor example, in this generation, people are watching movies and listening to musics more than
reading books and novels.
Input they are the boney and vital part o the society.
GPT-2-Small They are the most important part of the society.
GPT-2-Medium they are the vital part of the kingdom.
GPT-2-Large they are the vital and boney parts of the society.
GPT-2-XL they are the vital part, the boney part, and the society.
GPT-Neo-1.3B they are the vital part of society.
GPT-Neo-2.7B they are the boney and vital part of society.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
140
|
GPT-Neo-1.3B they are the vital part of society.
GPT-Neo-2.7B they are the boney and vital part of society.
GPT-J-6B they are the bone and vital part of the society.
Input And I am completely disagree with what the auther want to say.
GPT-2-Small And I am completely disagree with what the auther want to say.
GPT-2-Medium And I am entirely disagree with what the auther want to say.
GPT-2-Large And I am completely disagree with what the auther want to say
GPT-2-XL And I think the auther want to say completely differently.
GPT-Neo-1.3B And I disagree with what he say.
GPT-Neo-2.7B And I absolutely disagree with what the author want to say.
GPT-J-6B I completely disagree with what the author wants to say.
Table 23: Qualitative examples on JFLEG -clean. We manually highlight in blue and red phrases that we find to be associated
with the input and target styles, respectively.
25
Model Text
Input romeo, doff thy name; and for that name, which is no part of thee, take all myself
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
141
|
with the input and target styles, respectively.
25
Model Text
Input romeo, doff thy name; and for that name, which is no part of thee, take all myself
GPT-2-Small speak you, my dear nurse, to me...
GPT-2-Medium Romeo, Romeo, Romeo.
GPT-2-Large romo, doff thy name, take all myself, all i’ll say is, you’re a dirty old man.
GPT-2-XL roome, doff thy name; and for that name, which is no part of thee, take all myself.
GPT-Neo-1.3B romeo, get out of my sight; get out, i’ll not have it.
GPT-Neo-2.7B rommy, get rid of your name; and because your name, which is no part of you, needs to get rid of, take all of me.
GPT-J-6B romeo, take off your name, and for that name, which is not part of you, take all of me.
Input now art thou sociable, now art thou romeo; now art thou what thou art, by art as well as by nature.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
142
|
Input now art thou sociable, now art thou romeo; now art thou what thou art, by art as well as by nature.
GPT-2-Small now art thou romeo, now art thou romeo; now art thee what thou art, by art as well as by nature.
GPT-2-Medium what I have said in thy speech thou hast said enough of me, wherewith to destroy thy soul.
GPT-2-Large now art thou sociable, now art thou romeo; now art thou what thou art, by art as well as by nature.
GPT-2-XL now is thou sociable, now is thou Romeo; now is thou what thou art, by nature as well as by nature.
GPT-Neo-1.3B and now art thou what thou art, by art as well as by nature.
GPT-Neo-2.7B now are you sociable, now are you romeo; now are you what you are, by nature.
GPT-J-6B now you are friendly, now you are Romeo; now you are who you are, by nature as well as by art.
Input the sweetest honey is loathsome in his own deliciousness and in the taste confounds the appetite.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
143
|
Input the sweetest honey is loathsome in his own deliciousness and in the taste confounds the appetite.
GPT-2-Small the sweetest honey is loathsome in his own deliciousness and in the taste confounds the appetite.
GPT-2-Medium sweetest honey, sweetest bread, sweetest wine, sweetest dainties, do them to the soul’s delights.
GPT-2-Large the sweetest honey is loathsome in his own deliciousness; in the taste confounds the appetite.
GPT-2-XL the most sweet and clearest honey is loathsome in the taste and in the very smell confounds the appetite.
GPT-Neo-1.3Bthe honey is sweeter than ever it was in the honeycomb, and is almost enough to make one lose their senses and
die a corpse.
GPT-Neo-2.7B the sweetest milk is foul and sickly in his own deliciousness and in the taste confounds the appetite.
GPT-J-6B the sweetest honey is disgusting in its own deliciousness and in its taste it confounds the appetite.
Input unless philosophy can make a juliet, displant a town, reverse a prince’s doom, it helps not, it prevails not.
GPT-2-Small what am I thinking of?*
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
144
|
GPT-2-Small what am I thinking of?*
GPT-2-Mediumif philosophy can make a juliet, supplant your King, reverse his doom, reverse a prince’s doom, reverse a king’s
doom.
GPT-2-Large unless philosophy can make a juliet, displant a town, reverse a prince’s doom, it helps not, it prevails not.
GPT-2-XL unless it can convert a whore into a knight, it does not make it jolly.
GPT-Neo-1.3B unless philosophy can make a juliet, a town, reverse a prince’s doom, it helps not, it prevails not.
GPT-Neo-2.7Bunless philosophy can make a Juliet, cause a town to explode, reverse a prince’s downfall, it helps not, it prevails
not.
GPT-J-6B unless philosophy can make a Juliet, overturn a town, reverse a prince’s doom, it doesn’t help, it doesn’t work.
Input night’s candles are burnt out, and jocund day stands tiptoe on the misty mountain tops.
GPT-2-Small not here
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
145
|
Input night’s candles are burnt out, and jocund day stands tiptoe on the misty mountain tops.
GPT-2-Small not here
GPT-2-Medium and when night comes, the sun sets and stars all flicker in the distance.
GPT-2-Large night’s candles are burnt out, and jocund day stands in a field of snow.
GPT-2-XL night’s candles are burning out, and jocund day is standing still on the misty mountain tops.
GPT-Neo-1.3B night’s fires are out and the day stands up above the foggy mountains.
GPT-Neo-2.7B night’s candles are burnt out; day’s gone to bed and evening’s at hand; the sun is setting on the mountain tops.
GPT-J-6B day is over, and night is approaching, and it stands on its tiptoes over the misted mountain tops.
Input nor tears nor prayers shall purchase out abuses, therefore use none.
GPT-2-Small nor tears nor prayers shall purchase out abuses, therefore use none
GPT-2-Medium nor tears nor prayers shall purchase out abuses, therefore use none.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
146
|
GPT-2-Small nor tears nor prayers shall purchase out abuses, therefore use none
GPT-2-Medium nor tears nor prayers shall purchase out abuses, therefore use none.
GPT-2-Large use none, for abuse is in the eyes of a man.
GPT-2-XL no tears nor prayers shall buy out abuses, therefore I will not use any.
GPT-Neo-1.3B nor tears nor prayers shall have any power, therefore use none.
GPT-Neo-2.7B nor prayers nor tears will do any good, therefore use none.
GPT-J-6B and tears and prayers won’t work for you, so don’t use them.
Input and here stands all your state: either be gone before the watch be set, or by the break of day disguised from hence.
GPT-2-Small and here stands all your state: either be gone before the watch be set, or by the break of day disguised from hence.
GPT-2-Medium and this day there is nothing but the city.
GPT-2-Large but where is the watch?
GPT-2-XL and here stands all your state: either be gone before the watch, or by the break of day disguised from hence.
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
2205.11503
|
147
|
GPT-2-XL and here stands all your state: either be gone before the watch, or by the break of day disguised from hence.
GPT-Neo-1.3B and here stands all your state: either be gone before the watch be set, or by the break of day disguised from hence.
GPT-Neo-2.7B and this is all you get: either go before the watch is set, or before the break of day dressed like a thief from hence.
GPT-J-6B and here is all your stuff: either leave now or you’ll have to deal with us when it’s morning.
Table 24: Qualitative examples on SHAKESPEARE -clean. We manually highlight in blue and red phrases that we find to be
associated with the input and target styles, respectively. (Footnote*: Pray tell us, what are you thinking of right now?)
26
|
2205.11503
|
Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
|
We propose a method for arbitrary textual style transfer (TST)--the task of
transforming a text into any given style--utilizing general-purpose pre-trained
language models. Our method, Prompt-and-Rerank, is based on a mathematical
formulation of the TST task, decomposing it into three constituent components:
textual similarity, target style strength, and fluency. Specifically, our
method first uses zero-shot or few-shot prompting to obtain a set of candidate
generations in the target style, and then re-ranks these candidates according
to a combination of the three components above. Empirically, our method enables
small pre-trained language models to perform on par with state-of-the-art
large-scale models while consuming two orders of magnitude less compute and
memory. Finally, we conduct a systematic investigation of the effect of model
size and prompt design (e.g., prompt paraphrasing and delimiter-pair choice) on
style transfer quality across seven diverse textual style transfer datasets.
|
http://arxiv.org/pdf/2205.11503
|
[
"Mirac Suzgun",
"Luke Melas-Kyriazi",
"Dan Jurafsky"
] |
[
"cs.CL"
] |
GitHub page: https://github.com/suzgunmirac/prompt-and-rerank.
Project page: https://lukemelas.github.io/prompt-and-rerank/
| null |
cs.CL
|
20220523
|
20220523
|
[
{
"id": "2110.10668"
},
{
"id": "2105.08206"
},
{
"id": "1904.02295"
},
{
"id": "2112.07868"
},
{
"id": "1905.10060"
},
{
"id": "2205.11503"
},
{
"id": "2010.01054"
},
{
"id": "1908.08039"
}
] |
1907.11692
|
0
|
arXiv:1907.11692v1 [cs.CL] 26 Jul 2019RoBERTa: A Robustly Optimized BERT Pretraining Approach
Yinhan Liu∗§Myle Ott∗§Naman Goyal∗§Jingfei Du∗§Mandar Joshi†
Danqi Chen§Omer Levy§Mike Lewis§Luke Zettlemoyer†§Veselin Stoyanov§
†Paul G. Allen School of Computer Science & Engineering,
University of Washington, Seattle, WA
{mandar90,lsz }@cs.washington.edu
§Facebook AI
{yinhanliu,myleott,naman,jingfeidu,
danqi,omerlevy,mikelewis,lsz,ves }@fb.com
Abstract
Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different
sizes, and, as we will show, hyperparameter
choices have significant impact on the final results. We present a replication study of BERT
pretraining ( Devlin et al. ,2019 ) that carefully
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
1
|
pretraining ( Devlin et al. ,2019 ) that carefully
measures the impact of many key hyperparameters and training data size. We find that BERT
was significantly undertrained, and can match
or exceed the performance of every model
published after it. Our best model achieves
state-of-the-art results on GLUE, RACE and
SQuAD. These results highlight the importance of previously overlooked design choices,
and raise questions about the source of recently reported improvements. We release our
models and code.1
1 Introduction
Self-training methods such as ELMo ( Peters et al. ,
2018 ), GPT ( Radford et al. ,2018 ), BERT
(Devlin et al. ,2019 ), XLM ( Lample and Conneau ,
2019 ), and XLNet ( Yang et al. ,2019 ) have
brought significant performance gains, but it can
be challenging to determine which aspects of
the methods contribute the most. Training is
computationally expensive, limiting the amount
of tuning that can be done, and is often done with
private training data of varying sizes, limiting
our ability to measure the effects of the modeling
advances.
∗Equal contribution.
1Our models and code are available at:
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
2
|
private training data of varying sizes, limiting
our ability to measure the effects of the modeling
advances.
∗Equal contribution.
1Our models and code are available at:
https://github.com/pytorch/fairseqWe present a replication study of BERT pretraining ( Devlin et al. ,2019 ), which includes a
careful evaluation of the effects of hyperparmeter
tuning and training set size. We find that BERT
was significantly undertrained and propose an improved recipe for training BERT models, which
we call RoBERTa, that can match or exceed the
performance of all of the post-BERT methods.
Our modifications are simple, they include: (1)
training the model longer, with bigger batches,
over more data; (2) removing the next sentence
prediction objective; (3) training on longer sequences; and (4) dynamically changing the masking pattern applied to the training data. We also
collect a large new dataset (CC-N EWS) of comparable size to other privately used datasets, to better
control for training set size effects.
When controlling for training data, our improved training procedure improves upon the published BERT results on both GLUE and SQuAD.
When trained for longer over additional data, our
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
3
|
When controlling for training data, our improved training procedure improves upon the published BERT results on both GLUE and SQuAD.
When trained for longer over additional data, our
model achieves a score of 88.5 on the public
GLUE leaderboard, matching the 88.4 reported
byYang et al. (2019 ). Our model establishes a
new state-of-the-art on 4/9 of the GLUE tasks:
MNLI, QNLI, RTE and STS-B. We also match
state-of-the-art results on SQuAD and RACE.
Overall, we re-establish that BERT’s masked language model training objective is competitive
with other recently proposed training objectives
such as perturbed autoregressive language modeling (Yang et al. ,2019 ).2
In summary, the contributions of this paper
are: (1) We present a set of important BERT design choices and training strategies and introduce
2It is possible that these other methods could also improve
with more tuning. We leave this exploration to future work.
alternatives that lead to better downstream task
performance; (2) We use a novel dataset, CCNEWS, and confirm that using more data for pretraining further improves performance on downstream tasks; (3) Our training improvements show
that masked language model pretraining, under
the right design choices, is competitive with all
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
4
|
that masked language model pretraining, under
the right design choices, is competitive with all
other recently published methods. We release our
model, pretraining and fine-tuning code implemented in PyTorch ( Paszke et al. ,2017 ).
2 Background
In this section, we give a brief overview of the
BERT ( Devlin et al. ,2019 ) pretraining approach
and some of the training choices that we will examine experimentally in the following section.
2.1 Setup
BERT takes as input a concatenation of two
segments (sequences of tokens), x1,...,x N
andy1,...,yM. Segments usually consist of
more than one natural sentence. The two segments are presented as a single input sequence
to BERT with special tokens delimiting them:
[CLS],x1,...,x N,[SEP],y1,...,yM,[EOS].
MandNare constrained such that M+N < T ,
whereTis a parameter that controls the maximum
sequence length during training.
The model is first pretrained on a large unlabeled text corpus and subsequently finetuned using end-task labeled data.
2.2 Architecture
BERT uses the now ubiquitous transformer architecture ( Vaswani et al. ,2017 ), which we will not
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
5
|
2.2 Architecture
BERT uses the now ubiquitous transformer architecture ( Vaswani et al. ,2017 ), which we will not
review in detail. We use a transformer architecture
withLlayers. Each block uses Aself-attention
heads and hidden dimension H.
2.3 Training Objectives
During pretraining, BERT uses two objectives:
masked language modeling and next sentence prediction.
Masked Language Model (MLM) A random
sample of the tokens in the input sequence is
selected and replaced with the special token
[MASK]. The MLM objective is a cross-entropy
loss on predicting the masked tokens. BERT uniformly selects 15% of the input tokens for possible replacement. Of the selected tokens, 80% are
replaced with [MASK], 10% are left unchanged,and 10% are replaced by a randomly selected vocabulary token.
In the original implementation, random masking and replacement is performed once in the beginning and saved for the duration of training, although in practice, data is duplicated so the mask
is not always the same for every training sentence
(see Section 4.1).
Next Sentence Prediction (NSP) NSP is a binary classification loss for predicting whether two
segments follow each other in the original text.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
6
|
Next Sentence Prediction (NSP) NSP is a binary classification loss for predicting whether two
segments follow each other in the original text.
Positive examples are created by taking consecutive sentences from the text corpus. Negative examples are created by pairing segments from different documents. Positive and negative examples
are sampled with equal probability.
The NSP objective was designed to improve
performance on downstream tasks, such as Natural
Language Inference ( Bowman et al. ,2015 ), which
require reasoning about the relationships between
pairs of sentences.
2.4 Optimization
BERT is optimized with Adam ( Kingma and Ba ,
2015 ) using the following parameters: β1= 0.9,
β2= 0.999,ǫ=1e-6 and L2weight decay of0.01. The learning rate is warmed up
over the first 10,000 steps to a peak value of
1e-4, and then linearly decayed. BERT trains
with a dropout of 0.1 on all layers and attention weights, and a GELU activation function ( Hendrycks and Gimpel ,2016 ). Models are
pretrained for S=1,000,000 updates, with minibatches containing B=256 sequences of maximum length T=512 tokens.
2.5 Data
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
7
|
pretrained for S=1,000,000 updates, with minibatches containing B=256 sequences of maximum length T=512 tokens.
2.5 Data
BERT is trained on a combination of B OOK CORPUS (Zhu et al. ,2015 ) plus English W IKIPEDIA ,
which totals 16GB of uncompressed text.3
3 Experimental Setup
In this section, we describe the experimental setup
for our replication study of BERT.
3.1 Implementation
We reimplement BERT in FAIRSEQ (Ott et al. ,
2019 ). We primarily follow the original BERT
3Yang et al. (2019 ) use the same dataset but report having
only 13GB of text after data cleaning. This is most likely due
to subtle differences in cleaning of the Wikipedia data.
optimization hyperparameters, given in Section 2,
except for the peak learning rate and number of
warmup steps, which are tuned separately for each
setting. We additionally found training to be very
sensitive to the Adam epsilon term, and in some
cases we obtained better performance or improved
stability after tuning it. Similarly, we found setting
β2= 0.98to improve stability when training with
large batch sizes.
We pretrain with sequences of at most T= 512
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
8
|
β2= 0.98to improve stability when training with
large batch sizes.
We pretrain with sequences of at most T= 512
tokens. Unlike Devlin et al. (2019 ), we do not randomly inject short sequences, and we do not train
with a reduced sequence length for the first 90% of
updates. We train only with full-length sequences.
We train with mixed precision floating point
arithmetic on DGX-1 machines, each with 8 ×
32GB Nvidia V100 GPUs interconnected by Infiniband ( Micikevicius et al. ,2018 ).
3.2 Data
BERT-style pretraining crucially relies on large
quantities of text. Baevski et al. (2019 ) demonstrate that increasing data size can result in improved end-task performance. Several efforts
have trained on datasets larger and more diverse
than the original BERT ( Radford et al. ,2019 ;
Yang et al. ,2019 ;Zellers et al. ,2019 ). Unfortunately, not all of the additional datasets can be
publicly released. For our study, we focus on gathering as much data as possible for experimentation, allowing us to match the overall quality and
quantity of data as appropriate for each comparison.
We consider five English-language corpora of
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
9
|
quantity of data as appropriate for each comparison.
We consider five English-language corpora of
varying sizes and domains, totaling over 160GB
of uncompressed text. We use the following text
corpora:
•BOOK CORPUS (Zhu et al. ,2015 ) plus English
WIKIPEDIA . This is the original data used to
train BERT. (16GB).
•CC-N EWS, which we collected from the English portion of the CommonCrawl News
dataset ( Nagel ,2016 ). The data contains 63
million English news articles crawled between
September 2016 and February 2019. (76GB after filtering).4
•OPENWEBTEXT (Gokaslan and Cohen ,2019 ),
an open-source recreation of the WebText cor4We usenews-please (Hamborg et al. ,2017 ) to collect and extract CC-N EWS. CC-N EWS is similar to the R EALNEWS dataset described in Zellers et al. (2019 ).pus described in Radford et al. (2019 ). The text
is web content extracted from URLs shared on
Reddit with at least three upvotes. (38GB).5
•STORIES , a dataset introduced in Trinh and Le
(2018 ) containing a subset of CommonCrawl
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
10
|
Reddit with at least three upvotes. (38GB).5
•STORIES , a dataset introduced in Trinh and Le
(2018 ) containing a subset of CommonCrawl
data filtered to match the story-like style of
Winograd schemas. (31GB).
3.3 Evaluation
Following previous work, we evaluate our pretrained models on downstream tasks using the following three benchmarks.
GLUE The General Language Understanding Evaluation (GLUE) benchmark ( Wang et al. ,
2019b ) is a collection of 9 datasets for evaluating
natural language understanding systems.6Tasks
are framed as either single-sentence classification
or sentence-pair classification tasks. The GLUE
organizers provide training and development data
splits as well as a submission server and leaderboard that allows participants to evaluate and compare their systems on private held-out test data.
For the replication study in Section 4, we report
results on the development sets after finetuning
the pretrained models on the corresponding singletask training data (i.e., without multi-task training
or ensembling). Our finetuning procedure follows
the original BERT paper ( Devlin et al. ,2019 ).
In Section 5we additionally report test set results obtained from the public leaderboard. These
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
11
|
the original BERT paper ( Devlin et al. ,2019 ).
In Section 5we additionally report test set results obtained from the public leaderboard. These
results depend on a several task-specific modifications, which we describe in Section 5.1.
SQuAD The Stanford Question Answering
Dataset (SQuAD) provides a paragraph of context
and a question. The task is to answer the question
by extracting the relevant span from the context.
We evaluate on two versions of SQuAD: V1.1
and V2.0 ( Rajpurkar et al. ,2016 ,2018 ). In V1.1
the context always contains an answer, whereas in
5The authors and their affiliated institutions are not in any
way affiliated with the creation of the OpenWebText dataset.
6The datasets are: CoLA ( Warstadt et al. ,2018 ),
Stanford Sentiment Treebank (SST) ( Socher et al. ,
2013 ), Microsoft Research Paragraph Corpus
(MRPC) ( Dolan and Brockett ,2005 ), Semantic Textual Similarity Benchmark (STS) ( Agirre et al. ,2007 ),
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
12
|
(MRPC) ( Dolan and Brockett ,2005 ), Semantic Textual Similarity Benchmark (STS) ( Agirre et al. ,2007 ),
Quora Question Pairs (QQP) ( Iyer et al. ,2016 ), MultiGenre NLI (MNLI) ( Williams et al. ,2018 ), Question NLI
(QNLI) ( Rajpurkar et al. ,2016 ), Recognizing Textual
Entailment (RTE) ( Dagan et al. ,2006 ;Bar-Haim et al. ,
2006 ;Giampiccolo et al. ,2007 ;Bentivogli et al. ,2009 ) and
Winograd NLI (WNLI) ( Levesque et al. ,2011 ).
V2.0 some questions are not answered in the provided context, making the task more challenging.
For SQuAD V1.1 we adopt the same span prediction method as BERT ( Devlin et al. ,2019 ). For
SQuAD V2.0, we add an additional binary classifier to predict whether the question is answerable,
which we train jointly by summing the classification and span loss terms. During evaluation, we
only predict span indices on pairs that are classified as answerable.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
13
|
only predict span indices on pairs that are classified as answerable.
RACE The ReAding Comprehension from Examinations (RACE) ( Lai et al. ,2017 ) task is a
large-scale reading comprehension dataset with
more than 28,000 passages and nearly 100,000
questions. The dataset is collected from English
examinations in China, which are designed for
middle and high school students. In RACE, each
passage is associated with multiple questions. For
every question, the task is to select one correct answer from four options. RACE has significantly
longer context than other popular reading comprehension datasets and the proportion of questions
that requires reasoning is very large.
4 Training Procedure Analysis
This section explores and quantifies which choices
are important for successfully pretraining BERT
models. We keep the model architecture fixed.7
Specifically, we begin by training BERT models
with the same configuration as BERT BASE (L=
12,H= 768 ,A= 12 , 110M params).
4.1 Static vs. Dynamic Masking
As discussed in Section 2, BERT relies on randomly masking and predicting tokens. The original BERT implementation performed masking
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
14
|
4.1 Static vs. Dynamic Masking
As discussed in Section 2, BERT relies on randomly masking and predicting tokens. The original BERT implementation performed masking
once during data preprocessing, resulting in a singlestatic mask. To avoid using the same mask for
each training instance in every epoch, training data
was duplicated 10 times so that each sequence is
masked in 10 different ways over the 40 epochs of
training. Thus, each training sequence was seen
with the same mask four times during training.
We compare this strategy with dynamic maskingwhere we generate the masking pattern every
time we feed a sequence to the model. This becomes crucial when pretraining for more steps or
with larger datasets.
7Studying architectural changes, including larger architectures, is an important area for future work.Masking SQuAD 2.0 MNLI-m SST-2
reference 76.3 84.3 92.8
Our reimplementation:
static 78.3 84.3 92.5
dynamic 78.7 84.0 92.9
Table 1: Comparison between static and dynamic
masking for BERT BASE. We report F1 for SQuAD and
accuracy for MNLI-m and SST-2. Reported results are
medians over 5 random initializations (seeds). Reference results are from Yang et al. (2019 ).
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
15
|
accuracy for MNLI-m and SST-2. Reported results are
medians over 5 random initializations (seeds). Reference results are from Yang et al. (2019 ).
Results Table 1compares the published
BERT BASE results from Devlin et al. (2019 ) to our
reimplementation with either static or dynamic
masking. We find that our reimplementation
with static masking performs similar to the
original BERT model, and dynamic masking is
comparable or slightly better than static masking.
Given these results and the additional efficiency
benefits of dynamic masking, we use dynamic
masking in the remainder of the experiments.
4.2 Model Input Format and Next Sentence
Prediction
In the original BERT pretraining procedure, the
model observes two concatenated document segments, which are either sampled contiguously
from the same document (with p= 0.5) or from
distinct documents. In addition to the masked language modeling objective, the model is trained to
predict whether the observed document segments
come from the same or distinct documents via an
auxiliary Next Sentence Prediction (NSP) loss.
The NSP loss was hypothesized to be an important factor in training the original BERT model.
Devlin et al. (2019 ) observe that removing NSP
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
16
|
The NSP loss was hypothesized to be an important factor in training the original BERT model.
Devlin et al. (2019 ) observe that removing NSP
hurts performance, with significant performance
degradation on QNLI, MNLI, and SQuAD 1.1.
However, some recent work has questioned the
necessity of the NSP loss ( Lample and Conneau ,
2019 ;Yang et al. ,2019 ;Joshi et al. ,2019 ).
To better understand this discrepancy, we compare several alternative training formats:
•SEGMENT -PAIR +NSP: This follows the original
input format used in BERT ( Devlin et al. ,2019 ),
with the NSP loss. Each input has a pair of segments, which can each contain multiple natural
sentences, but the total combined length must
be less than 512 tokens.
Model SQuAD 1.1/2.0 MNLI-m SST-2 RACE
Our reimplementation (with NSP loss):
SEGMENT -PAIR 90.4/78.7 84.0 92.9 64.2
SENTENCE -PAIR 88.7/76.2 82.9 92.1 63.0
Our reimplementation (without NSP loss):
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
17
|
SENTENCE -PAIR 88.7/76.2 82.9 92.1 63.0
Our reimplementation (without NSP loss):
FULL -SENTENCES 90.4/79.1 84.7 92.5 64.8
DOC-SENTENCES 90.6/79.7 84.7 92.7 65.6
BERT BASE 88.5/76.3 84.3 92.8 64.3
XLNet BASE (K = 7) –/81.3 85.8 92.7 66.1
XLNet BASE (K = 6) –/81.0 85.6 93.4 66.7
Table 2: Development set results for base models pretrained over B OOK CORPUS and W IKIPEDIA . All models are
trained for 1M steps with a batch size of 256 sequences. We rep ort F1 for SQuAD and accuracy for MNLI-m,
SST-2 and RACE. Reported results are medians over five random initializations (seeds). Results for BERT BASEand
XLNet BASEare from Yang et al. (2019 ).
•SENTENCE -PAIR +NSP: Each input contains a
pair of natural sentences , either sampled from
a contiguous portion of one document or from
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
18
|
•SENTENCE -PAIR +NSP: Each input contains a
pair of natural sentences , either sampled from
a contiguous portion of one document or from
separate documents. Since these inputs are significantly shorter than 512 tokens, we increase
the batch size so that the total number of tokens
remains similar to SEGMENT -PAIR +NSP. We retain the NSP loss.
•FULL -SENTENCES : Each input is packed with
full sentences sampled contiguously from one
or more documents, such that the total length is
at most 512 tokens. Inputs may cross document
boundaries. When we reach the end of one document, we begin sampling sentences from the
next document and add an extra separator token
between documents. We remove the NSP loss.
•DOC-SENTENCES : Inputs are constructed similarly to FULL -SENTENCES , except that they
may not cross document boundaries. Inputs
sampled near the end of a document may be
shorter than 512 tokens, so we dynamically increase the batch size in these cases to achieve
a similar number of total tokens as FULL SENTENCES . We remove the NSP loss.
Results Table 2shows results for the four different settings. We first compare the original
SEGMENT -PAIR input format from Devlin et al.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
19
|
Results Table 2shows results for the four different settings. We first compare the original
SEGMENT -PAIR input format from Devlin et al.
(2019 ) to the SENTENCE -PAIR format; both formats retain the NSP loss, but the latter uses single sentences. We find that using individual
sentences hurts performance on downstream
tasks , which we hypothesize is because the model
is not able to learn long-range dependencies.We next compare training without the NSP
loss and training with blocks of text from a single document ( DOC-SENTENCES ). We find that
this setting outperforms the originally published
BERT BASEresults and that removing the NSP loss
matches or slightly improves downstream task
performance , in contrast to Devlin et al. (2019 ).
It is possible that the original BERT implementation may only have removed the loss term while
still retaining the SEGMENT -PAIR input format.
Finally we find that restricting sequences to
come from a single document ( DOC-SENTENCES )
performs slightly better than packing sequences
from multiple documents ( FULL -SENTENCES ).
However, because the DOC-SENTENCES format
results in variable batch sizes, we use FULL SENTENCES in the remainder of our experiments
for easier comparison with related work.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
20
|
However, because the DOC-SENTENCES format
results in variable batch sizes, we use FULL SENTENCES in the remainder of our experiments
for easier comparison with related work.
4.3 Training with large batches
Past work in Neural Machine Translation has
shown that training with very large mini-batches
can both improve optimization speed and end-task
performance when the learning rate is increased
appropriately ( Ott et al. ,2018 ). Recent work has
shown that BERT is also amenable to large batch
training ( You et al. ,2019 ).
Devlin et al. (2019 ) originally trained
BERT BASE for 1M steps with a batch size of
256 sequences. This is equivalent in computational cost, via gradient accumulation, to training
for 125K steps with a batch size of 2K sequences,
or for 31K steps with a batch size of 8K.
In Table 3we compare perplexity and endbsz steps lr ppl MNLI-m SST-2
256 1M 1e-4 3.99 84.7 92.7
2K 125K 7e-4 3.68 85.2 92.9
8K 31K 1e-3 3.77 84.6 92.8
Table 3: Perplexity on held-out training data ( ppl) and
development set accuracy for base models trained over
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
21
|
Table 3: Perplexity on held-out training data ( ppl) and
development set accuracy for base models trained over
BOOK CORPUS and W IKIPEDIA with varying batch
sizes ( bsz). We tune the learning rate ( lr) for each setting. Models make the same number of passes over the
data (epochs) and have the same computational cost.
task performance of BERT BASE as we increase the
batch size, controlling for the number of passes
through the training data. We observe that training with large batches improves perplexity for the
masked language modeling objective, as well as
end-task accuracy. Large batches are also easier to
parallelize via distributed data parallel training,8
and in later experiments we train with batches of
8K sequences.
Notably You et al. (2019 ) train BERT with even
larger batche sizes, up to 32K sequences. We leave
further exploration of the limits of large batch
training to future work.
4.4 Text Encoding
Byte-Pair Encoding (BPE) ( Sennrich et al. ,2016 )
is a hybrid between character- and word-level representations that allows handling the large vocabularies common in natural language corpora. Instead of full words, BPE relies on subwords units,
which are extracted by performing statistical analysis of the training corpus.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
22
|
which are extracted by performing statistical analysis of the training corpus.
BPE vocabulary sizes typically range from
10K-100K subword units. However, unicode characters can account for a sizeable portion of this
vocabulary when modeling large and diverse corpora, such as the ones considered in this work.
Radford et al. (2019 ) introduce a clever implementation of BPE that uses bytes instead of unicode characters as the base subword units. Using
bytes makes it possible to learn a subword vocabulary of a modest size (50K units) that can still encode any input text without introducing any “unknown” tokens.
8Large batch training can improve training efficiency even
without large scale parallel hardware through gradient accumulation , whereby gradients from multiple mini-batches
are accumulated locally before each optimization step. Thi s
functionality is supported natively in FAIRSEQ (Ott et al. ,
2019 ).The original BERT implementation ( Devlin et al. ,2019 ) uses a character-level
BPE vocabulary of size 30K, which is learned
after preprocessing the input with heuristic tokenization rules. Following Radford et al. (2019 ),
we instead consider training BERT with a larger
byte-level BPE vocabulary containing 50K subword units, without any additional preprocessing
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
23
|
we instead consider training BERT with a larger
byte-level BPE vocabulary containing 50K subword units, without any additional preprocessing
or tokenization of the input. This adds approximately 15M and 20M additional parameters for
BERT BASEand BERT LARGE , respectively.
Early experiments revealed only slight differences between these encodings, with the
Radford et al. (2019 ) BPE achieving slightly
worse end-task performance on some tasks. Nevertheless, we believe the advantages of a universal encoding scheme outweighs the minor degredation in performance and use this encoding in
the remainder of our experiments. A more detailed comparison of these encodings is left to future work.
5 RoBERTa
In the previous section we propose modifications
to the BERT pretraining procedure that improve
end-task performance. We now aggregate these
improvements and evaluate their combined impact. We call this configuration RoBERTa for
Robustly optimized BERT approach. Specifically, RoBERTa is trained with dynamic masking (Section 4.1),FULL -SENTENCES without NSP
loss (Section 4.2), large mini-batches (Section 4.3)
and a larger byte-level BPE (Section 4.4).
Additionally, we investigate two other important factors that have been under-emphasized in
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
24
|
and a larger byte-level BPE (Section 4.4).
Additionally, we investigate two other important factors that have been under-emphasized in
previous work: (1) the data used for pretraining,
and (2) the number of training passes through the
data. For example, the recently proposed XLNet
architecture ( Yang et al. ,2019 ) is pretrained using nearly 10 times more data than the original
BERT ( Devlin et al. ,2019 ). It is also trained with
a batch size eight times larger for half as many optimization steps, thus seeing four times as many
sequences in pretraining compared to BERT.
To help disentangle the importance of these factors from other modeling choices (e.g., the pretraining objective), we begin by training RoBERTa
following the BERT LARGE architecture ( L= 24 ,
H= 1024 ,A= 16 , 355M parameters). We
pretrain for 100K steps over a comparable B OOK CORPUS plus W IKIPEDIA dataset as was used in
Model data bsz stepsSQuADMNLI-m SST-2(v1.1/2.0)
RoBERTa
with B OOKS + W IKI 16GB 8K 100K 93.6/87.3 89.0 95.3
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
25
|
RoBERTa
with B OOKS + W IKI 16GB 8K 100K 93.6/87.3 89.0 95.3
+ additional data ( §3.2) 160GB 8K 100K 94.0/87.7 89.3 95.6
+ pretrain longer 160GB 8K 300K 94.4/88.7 90.0 96.1
+ pretrain even longer 160GB 8K 500K 94.6/89.4 90.2 96.4
BERT LARGE
with B OOKS + W IKI 13GB 256 1M 90.9/81.8 86.6 93.7
XLNet LARGE
with B OOKS + W IKI 13GB 256 1M 94.0/87.8 88.4 94.4
+ additional data 126GB 2K 500K 94.5/88.8 89.8 95.6
Table 4: Development set results for RoBERTa as we pretrain o ver more data (16GB →160GB of text) and pretrain
for longer (100K →300K→500K steps). Each row accumulates improvements from the row s above. RoBERTa
matches the architecture and training objective of BERT LARGE . Results for BERT LARGE and XLNet LARGE are from
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
26
|
matches the architecture and training objective of BERT LARGE . Results for BERT LARGE and XLNet LARGE are from
Devlin et al. (2019 ) and Yang et al. (2019 ), respectively. Complete results on all GLUE tasks can be fo und in the
Appendix.
Devlin et al. (2019 ). We pretrain our model using
1024 V100 GPUs for approximately one day.
Results We present our results in Table 4. When
controlling for training data, we observe that
RoBERTa provides a large improvement over the
originally reported BERT LARGE results, reaffirming
the importance of the design choices we explored
in Section 4.
Next, we combine this data with the three additional datasets described in Section 3.2. We
train RoBERTa over the combined data with the
same number of training steps as before (100K).
In total, we pretrain over 160GB of text. We observe further improvements in performance across
all downstream tasks, validating the importance of
data size and diversity in pretraining.9
Finally, we pretrain RoBERTa for significantly
longer, increasing the number of pretraining steps
from 100K to 300K, and then further to 500K. We
again observe significant gains in downstream task
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
27
|
longer, increasing the number of pretraining steps
from 100K to 300K, and then further to 500K. We
again observe significant gains in downstream task
performance, and the 300K and 500K step models outperform XLNet LARGE across most tasks. We
note that even our longest-trained model does not
appear to overfit our data and would likely benefit
from additional training.
In the rest of the paper, we evaluate our best
RoBERTa model on the three different benchmarks: GLUE, SQuaD and RACE. Specifically
9Our experiments conflate increases in data size and diversity. We leave a more careful analysis of these two dimensions to future work.we consider RoBERTa trained for 500K steps over
all five of the datasets introduced in Section 3.2.
5.1 GLUE Results
For GLUE we consider two finetuning settings.
In the first setting ( single-task, dev ) we finetune
RoBERTa separately for each of the GLUE tasks,
using only the training data for the corresponding task. We consider a limited hyperparameter
sweep for each task, with batch sizes ∈ {16,32}
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
28
|
using only the training data for the corresponding task. We consider a limited hyperparameter
sweep for each task, with batch sizes ∈ {16,32}
and learning rates ∈ {1e−5,2e−5,3e−5}, with a
linear warmup for the first 6% of steps followed by
a linear decay to 0. We finetune for 10 epochs and
perform early stopping based on each task’s evaluation metric on the dev set. The rest of the hyperparameters remain the same as during pretraining.
In this setting, we report the median development
set results for each task over five random initializations, without model ensembling.
In the second setting ( ensembles, test ), we compare RoBERTa to other approaches on the test set
via the GLUE leaderboard. While many submissions to the GLUE leaderboard depend on multitask finetuning, our submission depends only on
single-task finetuning . For RTE, STS and MRPC
we found it helpful to finetune starting from the
MNLI single-task model, rather than the baseline
pretrained RoBERTa. We explore a slightly wider
hyperparameter space, described in the Appendix,
and ensemble between 5 and 7 models per task.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
29
|
pretrained RoBERTa. We explore a slightly wider
hyperparameter space, described in the Appendix,
and ensemble between 5 and 7 models per task.
MNLI QNLI QQP RTE SST MRPC CoLA STS WNLI Avg
Single-task single models on dev
BERT LARGE 86.6/- 92.3 91.3 70.4 93.2 88.0 60.6 90.0 - XLNet LARGE 89.8/- 93.9 91.8 83.8 95.6 89.2 63.6 91.8 - RoBERTa 90.2/90.2 94.7 92.2 86.6 96.4 90.9 68.0 92.4 91.3 Ensembles on test (from leaderboard as of July 25, 2019)
ALICE 88.2/87.9 95.7 90.7 83.5 95.2 92.6 68.6 91.1 80.8 86.3
MT-DNN 87.9/87.4 96.0 89.9 86.3 96.5 92.7 68.4 91.1 89.0 87.6
XLNet 90.2/89.8 98.6 90.3 86.3 96.8 93.0 67.8 91.6 90.4 88.4
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
30
|
XLNet 90.2/89.8 98.6 90.3 86.3 96.8 93.0 67.8 91.6 90.4 88.4
RoBERTa 90.8/90.2 98.9 90.2 88.2 96.7 92.3 67.8 92.2 89.0 88.5
Table 5: Results on GLUE. All results are based on a 24-layer a rchitecture. BERT LARGE and XLNet LARGE results
are from Devlin et al. (2019 ) and Yang et al. (2019 ), respectively. RoBERTa results on the development set are a
median over five runs. RoBERTa results on the test set are ense mbles of single-task models. For RTE, STS and
MRPC we finetune starting from the MNLI model instead of the ba seline pretrained model. Averages are obtained
from the GLUE leaderboard.
Task-specific modifications Two of the GLUE
tasks require task-specific finetuning approaches
to achieve competitive leaderboard results.
QNLI : Recent submissions on the GLUE
leaderboard adopt a pairwise ranking formulation
for the QNLI task, in which candidate answers
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
31
|
to achieve competitive leaderboard results.
QNLI : Recent submissions on the GLUE
leaderboard adopt a pairwise ranking formulation
for the QNLI task, in which candidate answers
are mined from the training set and compared to
one another, and a single (question, candidate)
pair is classified as positive ( Liu et al. ,2019b ,a;
Yang et al. ,2019 ). This formulation significantly
simplifies the task, but is not directly comparable
to BERT ( Devlin et al. ,2019 ). Following recent
work, we adopt the ranking approach for our test
submission, but for direct comparison with BERT
we report development set results based on a pure
classification approach.
WNLI : We found the provided NLI-format
data to be challenging to work with. Instead
we use the reformatted WNLI data from SuperGLUE ( Wang et al. ,2019a ), which indicates the
span of the query pronoun and referent. We finetune RoBERTa using the margin ranking loss from
Kocijan et al. (2019 ). For a given input sentence,
we use spaCy ( Honnibal and Montani ,2017 ) to
extract additional candidate noun phrases from the
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
32
|
Kocijan et al. (2019 ). For a given input sentence,
we use spaCy ( Honnibal and Montani ,2017 ) to
extract additional candidate noun phrases from the
sentence and finetune our model so that it assigns
higher scores to positive referent phrases than for
any of the generated negative candidate phrases.
One unfortunate consequence of this formulation
is that we can only make use of the positive training examples, which excludes over half of the provided training examples.10
10While we only use the provided WNLI training data, ourResults We present our results in Table 5. In the
first setting ( single-task, dev ), RoBERTa achieves
state-of-the-art results on all 9 of the GLUE
task development sets. Crucially, RoBERTa uses
the same masked language modeling pretraining objective and architecture as BERT LARGE , yet
consistently outperforms both BERT LARGE and
XLNet LARGE . This raises questions about the relative importance of model architecture and pretraining objective, compared to more mundane details like dataset size and training time that we explore in this work.
In the second setting ( ensembles, test ), we
submit RoBERTa to the GLUE leaderboard and
achieve state-of-the-art results on 4 out of 9 tasks
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
33
|
submit RoBERTa to the GLUE leaderboard and
achieve state-of-the-art results on 4 out of 9 tasks
and the highest average score to date. This is especially exciting because RoBERTa does not depend
on multi-task finetuning, unlike most of the other
top submissions. We expect future work may further improve these results by incorporating more
sophisticated multi-task finetuning procedures.
5.2 SQuAD Results
We adopt a much simpler approach for SQuAD
compared to past work. In particular, while
both BERT ( Devlin et al. ,2019 ) and XLNet ( Yang et al. ,2019 ) augment their training data
with additional QA datasets, we only finetune
RoBERTa using the provided SQuAD training
data .Yang et al. (2019 ) also employed a custom
layer-wise learning rate schedule to finetune
results could potentially be improved by augmenting this wi th
additional pronoun disambiguation datasets.
ModelSQuAD 1.1 SQuAD 2.0
EM F1 EM F1
Single models on dev, w/o data augmentation
BERT LARGE 84.1 90.9 79.0 81.8
XLNet LARGE 89.0 94.5 86.1 88.8
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
34
|
BERT LARGE 84.1 90.9 79.0 81.8
XLNet LARGE 89.0 94.5 86.1 88.8
RoBERTa 88.9 94.6 86.5 89.4
Single models on test (as of July 25, 2019)
XLNet LARGE 86.3†89.1†
RoBERTa 86.8 89.8
XLNet + SG-Net Verifier 87.0†89.9†
Table 6: Results on SQuAD. †indicates results that depend on additional external training data. RoBERTa
uses only the provided SQuAD data in both dev and
test settings. BERT LARGE and XLNet LARGE results are
from Devlin et al. (2019 ) and Yang et al. (2019 ), respectively.
XLNet, while we use the same learning rate for
all layers.
For SQuAD v1.1 we follow the same finetuning procedure as Devlin et al. (2019 ). For SQuAD
v2.0, we additionally classify whether a given
question is answerable; we train this classifier
jointly with the span predictor by summing the
classification and span loss terms.
Results We present our results in Table 6. On
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
35
|
jointly with the span predictor by summing the
classification and span loss terms.
Results We present our results in Table 6. On
the SQuAD v1.1 development set, RoBERTa
matches the state-of-the-art set by XLNet. On the
SQuAD v2.0 development set, RoBERTa sets a
new state-of-the-art, improving over XLNet by 0.4
points (EM) and 0.6 points (F1).
We also submit RoBERTa to the public SQuAD
2.0 leaderboard and evaluate its performance relative to other systems. Most of the top systems
build upon either BERT ( Devlin et al. ,2019 ) or
XLNet ( Yang et al. ,2019 ), both of which rely on
additional external training data. In contrast, our
submission does not use any additional data.
Our single RoBERTa model outperforms all but
one of the single model submissions, and is the
top scoring system among those that do not rely
on data augmentation.
5.3 RACE Results
In RACE, systems are provided with a passage of
text, an associated question, and four candidate answers. Systems are required to classify which of
the four candidate answers is correct.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
36
|
In RACE, systems are provided with a passage of
text, an associated question, and four candidate answers. Systems are required to classify which of
the four candidate answers is correct.
We modify RoBERTa for this task by concate-Model Accuracy Middle High
Single models on test (as of July 25, 2019)
BERT LARGE 72.0 76.6 70.1
XLNet LARGE 81.7 85.4 80.2
RoBERTa 83.2 86.5 81.3
Table 7: Results on the RACE test set. BERT LARGE and
XLNet LARGE results are from Yang et al. (2019 ).
nating each candidate answer with the corresponding question and passage. We then encode each of
these four sequences and pass the resulting [CLS]
representations through a fully-connected layer,
which is used to predict the correct answer. We
truncate question-answer pairs that are longer than
128 tokens and, if needed, the passage so that the
total length is at most 512 tokens.
Results on the RACE test sets are presented in
Table 7. RoBERTa achieves state-of-the-art results
on both middle-school and high-school settings.
6 Related Work
Pretraining methods have been designed
with different training objectives, including language modeling ( Dai and Le ,2015 ;
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
37
|
on both middle-school and high-school settings.
6 Related Work
Pretraining methods have been designed
with different training objectives, including language modeling ( Dai and Le ,2015 ;
Peters et al. ,2018 ;Howard and Ruder ,2018 ),
machine translation ( McCann et al. ,2017 ), and
masked language modeling ( Devlin et al. ,2019 ;
Lample and Conneau ,2019 ). Many recent
papers have used a basic recipe of finetuning
models for each end task ( Howard and Ruder ,
2018 ;Radford et al. ,2018 ), and pretraining
with some variant of a masked language model
objective. However, newer methods have
improved performance by multi-task fine tuning ( Dong et al. ,2019 ), incorporating entity
embeddings ( Sun et al. ,2019 ), span prediction ( Joshi et al. ,2019 ), and multiple variants
of autoregressive pretraining ( Song et al. ,2019 ;
Chan et al. ,2019 ;Yang et al. ,2019 ). Performance is also typically improved by training
bigger models on more data ( Devlin et al. ,
2019 ;Baevski et al. ,2019 ;Yang et al. ,2019 ;
Radford et al. ,2019 ). Our goal was to replicate,
simplify, and better tune the training of BERT,
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
38
|
Radford et al. ,2019 ). Our goal was to replicate,
simplify, and better tune the training of BERT,
as a reference point for better understanding the
relative performance of all of these methods.
7 Conclusion
We carefully evaluate a number of design decisions when pretraining BERT models. We
find that performance can be substantially improved by training the model longer, with bigger
batches over more data; removing the next sentence prediction objective; training on longer sequences; and dynamically changing the masking
pattern applied to the training data. Our improved
pretraining procedure, which we call RoBERTa,
achieves state-of-the-art results on GLUE, RACE
and SQuAD, without multi-task finetuning for
GLUE or additional data for SQuAD. These results illustrate the importance of these previously overlooked design decisions and suggest
that BERT’s pretraining objective remains competitive with recently proposed alternatives.
We additionally use a novel dataset,
CC-N EWS, and release our models and
code for pretraining and finetuning at:
https://github.com/pytorch/fairseq .
References
Eneko Agirre, Llu’is M‘arquez, and Richard Wicentowski, editors. 2007. Proceedings of the Fourth
International Workshop on Semantic Evaluations
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
39
|
Eneko Agirre, Llu’is M‘arquez, and Richard Wicentowski, editors. 2007. Proceedings of the Fourth
International Workshop on Semantic Evaluations
(SemEval-2007) .
Alexei Baevski, Sergey Edunov, Yinhan Liu, Luke
Zettlemoyer, and Michael Auli. 2019. Clozedriven pretraining of self-attention networks. arXiv
preprint arXiv:1903.07785 .
Roy Bar-Haim, Ido Dagan, Bill Dolan, Lisa Ferro,
Danilo Giampiccolo, Bernardo Magnini, and Idan
Szpektor. 2006. The second PASCAL recognising
textual entailment challenge. In Proceedings of the
second PASCAL challenges workshop on recognising textual entailment .
Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo
Giampiccolo, and Bernardo Magnini. 2009. The
fifth PASCAL recognizing textual entailment challenge.
Samuel R Bowman, Gabor Angeli, Christopher Potts,
and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference.
InEmpirical Methods in Natural Language Processing (EMNLP) .
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
40
|
and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference.
InEmpirical Methods in Natural Language Processing (EMNLP) .
William Chan, Nikita Kitaev, Kelvin Guu, Mitchell
Stern, and Jakob Uszkoreit. 2019. KERMIT: Generative insertion-based modeling for sequences. arXiv
preprint arXiv:1906.01604 .Ido Dagan, Oren Glickman, and Bernardo Magnini.
2006. The PASCAL recognising textual entailment
challenge. In Machine learning challenges. evaluating predictive uncertainty, visual object classification, and recognising tectual entailment .
Andrew M Dai and Quoc V Le. 2015. Semi-supervised
sequence learning. In Advances in Neural Information Processing Systems (NIPS) .
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training of
deep bidirectional transformers for language understanding. In North American Association for Computational Linguistics (NAACL) .
William B Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the International Workshop on Paraphrasing .
Li Dong, Nan Yang, Wenhui Wang, Furu Wei,
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
41
|
Li Dong, Nan Yang, Wenhui Wang, Furu Wei,
Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming
Zhou, and Hsiao-Wuen Hon. 2019. Unified
language model pre-training for natural language
understanding and generation. arXiv preprint
arXiv:1905.03197 .
Danilo Giampiccolo, Bernardo Magnini, Ido Dagan,
and Bill Dolan. 2007. The third PASCAL recognizing textual entailment challenge. In Proceedings
of the ACL-PASCAL workshop on textual entailment
and paraphrasing .
Aaron Gokaslan and Vanya Cohen. 2019. Openwebtext corpus. http://web.archive.org/
save/http://Skylion007.github.io/
OpenWebTextCorpus .
Felix Hamborg, Norman Meuschke, Corinna Breitinger, and Bela Gipp. 2017. news-please: A
generic news crawler and extractor. In Proceedings
of the 15th International Symposium of Information
Science .
Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). arXiv preprint
arXiv:1606.08415 .
Matthew Honnibal and Ines Montani. 2017. spaCy 2:
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
42
|
arXiv:1606.08415 .
Matthew Honnibal and Ines Montani. 2017. spaCy 2:
Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear.
Jeremy Howard and Sebastian Ruder. 2018. Universal
language model fine-tuning for text classification.
arXiv preprint arXiv:1801.06146 .
Shankar Iyer, Nikhil Dandekar, and Kornl Csernai. 2016. First quora dataset release: Question
pairs.https://data.quora.com/FirstQuora-Dataset-Release-QuestionPairs .
Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S.
Weld, Luke Zettlemoyer, and Omer Levy. 2019.
SpanBERT: Improving pre-training by representing and predicting spans. arXiv preprint
arXiv:1907.10529 .
Diederik Kingma and Jimmy Ba. 2015. Adam: A
method for stochastic optimization. In International
Conference on Learning Representations (ICLR) .
Vid Kocijan, Ana-Maria Cretu, Oana-Maria Camburu,
Yordan Yordanov, and Thomas Lukasiewicz. 2019.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
43
|
Vid Kocijan, Ana-Maria Cretu, Oana-Maria Camburu,
Yordan Yordanov, and Thomas Lukasiewicz. 2019.
A surprisingly robust trick for winograd schema
challenge. arXiv preprint arXiv:1905.06290 .
Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang,
and Eduard Hovy. 2017. Race: Large-scale reading
comprehension dataset from examinations. arXiv
preprint arXiv:1704.04683 .
Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint
arXiv:1901.07291 .
Hector J Levesque, Ernest Davis, and Leora Morgenstern. 2011. The Winograd schema challenge. In
AAAI Spring Symposium: Logical Formalizations of
Commonsense Reasoning .
Xiaodong Liu, Pengcheng He, Weizhu Chen, and
Jianfeng Gao. 2019a. Improving multi-task deep
neural networks via knowledge distillation for
natural language understanding. arXiv preprint
arXiv:1904.09482 .
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
44
|
neural networks via knowledge distillation for
natural language understanding. arXiv preprint
arXiv:1904.09482 .
Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. 2019b. Multi-task deep neural networks
for natural language understanding. arXiv preprint
arXiv:1901.11504 .
Bryan McCann, James Bradbury, Caiming Xiong, and
Richard Socher. 2017. Learned in translation: Contextualized word vectors. In Advances in Neural Information Processing Systems (NIPS) , pages 6297–
6308.
Paulius Micikevicius, Sharan Narang, Jonah Alben,
Gregory Diamos, Erich Elsen, David Garcia, Boris
Ginsburg, Michael Houston, Oleksii Kuchaiev,
Ganesh Venkatesh, and Hao Wu. 2018. Mixed precision training. In International Conference on Learning Representations .
Sebastian Nagel. 2016. Cc-news. http:
//web.archive.org/save/http:
//commoncrawl.org/2016/10/newsdataset-available .
Myle Ott, Sergey Edunov, Alexei Baevski, Angela
Fan, Sam Gross, Nathan Ng, David Grangier, and
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
45
|
Myle Ott, Sergey Edunov, Alexei Baevski, Angela
Fan, Sam Gross, Nathan Ng, David Grangier, and
Michael Auli. 2019. FAIRSEQ : A fast, extensible toolkit for sequence modeling. In North
American Association for Computational Linguistics (NAACL): System Demonstrations .Myle Ott, Sergey Edunov, David Grangier, and
Michael Auli. 2018. Scaling neural machine translation. In Proceedings of the Third Conference on
Machine Translation (WMT) .
Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam
Lerer. 2017. Automatic differentiation in PyTorch.
InNIPS Autodiff Workshop .
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word representations. In North American Association for Computational Linguistics (NAACL) .
Alec Radford, Karthik Narasimhan, Time Salimans,
and Ilya Sutskever. 2018. Improving language understanding with unsupervised learning. Technical
report, OpenAI.
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
46
|
and Ilya Sutskever. 2018. Improving language understanding with unsupervised learning. Technical
report, OpenAI.
Alec Radford, Jeffrey Wu, Rewon Child, David Luan,
Dario Amodei, and Ilya Sutskever. 2019. Language
models are unsupervised multitask learners. Technical report, OpenAI.
Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018.
Know what you don’t know: Unanswerable questions for squad. In Association for Computational
Linguistics (ACL) .
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
Percy Liang. 2016. SQuAD: 100,000+ questions for
machine comprehension of text. In Empirical Methods in Natural Language Processing (EMNLP) .
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016. Neural machine translation of rare words with
subword units. In Association for Computational
Linguistics (ACL) , pages 1715–1725.
Richard Socher, Alex Perelygin, Jean Wu, Jason
Chuang, Christopher D Manning, Andrew Ng, and
Christopher Potts. 2013. Recursive deep models
for semantic compositionality over a sentiment treebank. In Empirical Methods in Natural Language
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
47
|
Christopher Potts. 2013. Recursive deep models
for semantic compositionality over a sentiment treebank. In Empirical Methods in Natural Language
Processing (EMNLP) .
Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and
Tie-Yan Liu. 2019. MASS: Masked sequence
to sequence pre-training for language generation.
InInternational Conference on Machine Learning
(ICML) .
Yu Stephanie Sun, Shuohuan Wang, Yukun Li, Shikun
Feng, Xuyi Chen, Han Zhang, Xinlun Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. ERNIE: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223 .
Trieu H Trinh and Quoc V Le. 2018. A simple
method for commonsense reasoning. arXiv preprint
arXiv:1806.02847 .
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. In Advances in neural information processing systems .
Alex Wang, Yada Pruksachatkun, Nikita Nangia,
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
1907.11692
|
48
|
you need. In Advances in neural information processing systems .
Alex Wang, Yada Pruksachatkun, Nikita Nangia,
Amanpreet Singh, Julian Michael, Felix Hill, Omer
Levy, and Samuel R. Bowman. 2019a. SuperGLUE:
A stickier benchmark for general-purpose language
understanding systems. arXiv preprint 1905.00537 .
Alex Wang, Amanpreet Singh, Julian Michael, Felix
Hill, Omer Levy, and Samuel R. Bowman. 2019b.
GLUE: A multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations
(ICLR) .
Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2018. Neural network acceptability judgments. arXiv preprint 1805.12471 .
Adina Williams, Nikita Nangia, and Samuel Bowman.
2018. A broad-coverage challenge corpus for sentence understanding through inference. In North
American Association for Computational Linguistics (NAACL) .
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le.
2019. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint
arXiv:1906.08237 .
|
1907.11692
|
RoBERTa: A Robustly Optimized BERT Pretraining Approach
|
Language model pretraining has led to significant performance gains but
careful comparison between different approaches is challenging. Training is
computationally expensive, often done on private datasets of different sizes,
and, as we will show, hyperparameter choices have significant impact on the
final results. We present a replication study of BERT pretraining (Devlin et
al., 2019) that carefully measures the impact of many key hyperparameters and
training data size. We find that BERT was significantly undertrained, and can
match or exceed the performance of every model published after it. Our best
model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results
highlight the importance of previously overlooked design choices, and raise
questions about the source of recently reported improvements. We release our
models and code.
|
http://arxiv.org/pdf/1907.11692
|
[
"Yinhan Liu",
"Myle Ott",
"Naman Goyal",
"Jingfei Du",
"Mandar Joshi",
"Danqi Chen",
"Omer Levy",
"Mike Lewis",
"Luke Zettlemoyer",
"Veselin Stoyanov"
] |
[
"cs.CL"
] | null | null |
cs.CL
|
20190726
|
20190726
|
[
{
"id": "1906.01604"
},
{
"id": "1904.09223"
},
{
"id": "1806.02847"
},
{
"id": "1704.04683"
},
{
"id": "1905.06290"
},
{
"id": "1506.06724"
},
{
"id": "1904.00962"
},
{
"id": "1606.08415"
},
{
"id": "1901.07291"
},
{
"id": "1907.11692"
},
{
"id": "1906.08237"
},
{
"id": "1801.06146"
},
{
"id": "1905.03197"
},
{
"id": "1907.10529"
},
{
"id": "1904.09482"
},
{
"id": "1905.12616"
},
{
"id": "1901.11504"
},
{
"id": "1903.07785"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.