sinhala-nlp/sinhala-headline-prediction

Sinhala Headline Prediction

Sinhala Headline Prediction is a binary text classification dataset for Sinhala introduced as part of the SINHALA-GLUE benchmark in the ACL 2025 paper "Sinhala Encoder-only Language Models and Evaluation".

Given a Sinhala news article and a candidate headline, the task is to determine whether the headline is the correct headline for the article.

The dataset contains 7,870 training examples and 1,970 test examples.

Dataset Description

News Headline Prediction (NHP) was introduced as a new task in SINHALA-GLUE to evaluate whether a language model can identify an appropriate headline for a Sinhala news article.

The dataset was constructed using Sinhala news articles and their corresponding headlines from NSina, a large Sinhala news corpus.

For each example, a news article is paired with either:

• its original headline; or
• an incorrect headline taken from another article.

Incorrect headlines were selected based on substantial word overlap with the original article. This was done to avoid constructing overly simple negative examples using completely unrelated headlines.

The task is formulated as binary classification:

• 1 — the candidate is the correct headline for the article.
• 0 — the candidate is not the correct headline for the article.

Dataset Structure

Each example contains the following fields:

id            : string
news_content  : string
headline      : string
is_headline   : int

Fields

id
A unique identifier for the example.

news_content
The content of the Sinhala news article.

headline
A candidate headline associated with the article.

is_headline
The binary target label:

1 = correct headline
0 = incorrect headline

Data Splits

Split	Examples
Train	7,870
Test	1,970
Total	9,840

The train and test splits follow those used in the SINHALA-GLUE benchmark.

Example

An example has the following structure:

{
    "id": "hp_train_0001",
    "news_content": "...",
    "headline": "...",
    "is_headline": 1
}

A positive example indicates that the provided headline belongs to the corresponding news article. For a negative example, the candidate headline comes from another article but was selected to have lexical overlap with the article.

Loading the Dataset

The dataset can be loaded using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset(
    "sinhala-nlp/sinhala-headline-prediction"
)

print(dataset)

Access an example using:

print(dataset["train"][0])

The article, headline and label can be accessed using:

article = dataset["train"][0]["news_content"]
headline = dataset["train"][0]["headline"]
label = dataset["train"][0]["is_headline"]

Task

The task takes a news article and candidate headline as input:

[CLS] news_content [SEP] headline

and predicts whether the candidate is the correct headline:

0 or 1

The primary evaluation metric used for this task in SINHALA-GLUE is Macro F1.

Intended Uses

The dataset is primarily intended for research on Sinhala natural language processing, including:

• Sinhala language model evaluation;
• news headline prediction;
• text-pair classification;
• semantic matching between articles and headlines;
• representation learning;
• evaluation of monolingual and multilingual language models;
• research on low-resource NLP.

It forms one of the classification tasks in the SINHALA-GLUE benchmark.

Source Data

The underlying news articles and headlines originate from NSina, a Sinhala news corpus containing articles collected from ten Sinhala news websites in Sri Lanka.

The News Headline Prediction dataset was subsequently constructed from these article-headline pairs specifically for SINHALA-GLUE.

Dataset Creation

For positive examples, an article is paired with its original headline.

For negative examples, an article is paired with an incorrect headline. Instead of selecting headlines completely at random, candidate negative headlines were chosen based on substantial word overlap with the article.

This makes the negative examples more closely related to the corresponding articles and reduces the likelihood that the task can be solved simply by identifying completely unrelated headlines.

Languages

The dataset is primarily in Sinhala (si) and uses the Sinhala script.

Limitations

The dataset is derived from Sinhala news articles and therefore primarily represents the language and writing styles found in online news media.

Performance on this task should not be interpreted as a general measure of Sinhala language understanding.

The construction of negative examples is based partly on lexical overlap. Consequently, the resulting classification task may contain artefacts associated with the negative-sampling procedure.

The dataset also inherits potential biases, errors and coverage limitations from the underlying NSina news corpus.

Citation

If you use this dataset, please cite the paper that introduced the News Headline Prediction task and SINHALA-GLUE:

@inproceedings{ranasinghe-etal-2025-sinhala,
    title = "{S}inhala Encoder-only Language Models and Evaluation",
    author = "Ranasinghe, Tharindu  and
      Hettiarachchi, Hansi  and
      Pathirana, Nadeesha  and
      Premasiri, Damith  and
      Uyangodage, Lasitha  and
      Arachchige, Isuri  and
      Plum, Alistair  and
      Rayson, Paul  and
      Mitkov, Ruslan",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.422/",
    doi = "10.18653/v1/2025.acl-long.422",
    pages = "8623--8636",
    ISBN = "979-8-89176-251-0"
}

Related Resources

This dataset is one of the six tasks included in SINHALA-GLUE, the first natural language understanding benchmark developed specifically for Sinhala.

The benchmark includes:

• Sentiment Analysis (SA)
• Offensive Language Detection (OLD)
• News Headline Prediction (NHP)
• Semantic Textual Similarity (STS)
• Named Entity Recognition (NER)
• Offensive Token Detection (OTD)

SINHALA-GLUE and the associated Sinhala encoder models were introduced in "Sinhala Encoder-only Language Models and Evaluation", ACL 2025.