sinhala-nlp/named-entity-recognition

Sinhala Named Entity Recognition

Sinhala Named Entity Recognition is a token-level named entity recognition dataset for Sinhala.

This repository is a re-upload of the original Sinhala NER dataset introduced by Manamini et al. (2016) in "Ananya - a Named-Entity-Recognition (NER) System for Sinhala Language" with proper train/ test splits. The dataset was subsequently included as the Named Entity Recognition (NER) task in the SINHALA-GLUE benchmark introduced in "Sinhala Encoder-only Language Models and Evaluation" (Ranasinghe et al., 2025).

The dataset contains Sinhala news sentences annotated with named entities belonging to three categories:

• PER — Person
• ORG — Organisation
• LOC — Location
• O — Tokens that do not belong to a named entity

Dataset Description

The original dataset was developed for Sinhala named entity recognition by Manamini et al. (2016). The authors introduced the Ananya Sinhala NER system and used the annotated corpus to investigate named entity recognition for Sinhala.

The sentences in the dataset originate from Sinhala news articles and contain annotations for persons, organisations and locations.

This Hugging Face repository re-uploads the dataset in a format suitable for modern NLP libraries and was used as the NER task in the SINHALA-GLUE benchmark.

Dataset Structure

Each example contains the following fields:

id        : string
tokens    : list[string]
ner_tags  : list[string]

Fields

id
A unique identifier for the example.

tokens
A tokenised Sinhala sentence.

ner_tags
The named entity label corresponding to each token.

The label set is:

PER = Person
ORG = Organisation
LOC = Location
O   = Outside a named entity

Each token has a corresponding label at the same position in ner_tags.

Data Splits

Split	Examples
Train	4,000
Test	1,000
Total	5,000

These are the splits used for the NER task in SINHALA-GLUE.

Example

An example follows this structure:

{
    "id": "ner_train_0001",
    "tokens": [
        "ඒ",
        "කියන්නේ",
        "ලෝකයද",
        "නැවතත්",
        "සතුටු",
        "වෙන්නේය",
        "."
    ],
    "ner_tags": [
        "O",
        "O",
        "O",
        "O",
        "O",
        "O",
        "O"
    ]
}

For a sentence containing named entities, the corresponding tokens are labelled as PER, ORG, or LOC.

Loading the Dataset

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

from datasets import load_dataset

dataset = load_dataset(
    "sinhala-nlp/named-entity-recognition"
)

print(dataset)

An individual example can be accessed using:

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

The tokens and labels can be accessed using:

tokens = dataset["train"][0]["tokens"]
labels = dataset["train"][0]["ner_tags"]

Task

The task is formulated as token classification.

Given a sequence of Sinhala tokens, a model predicts one named entity label for each token:

Sinhala sentence
      ↓
PER / ORG / LOC / O

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

Intended Uses

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

• named entity recognition;
• token classification;
• Sinhala language model evaluation;
• sequence labelling;
• representation learning;
• evaluation of monolingual and multilingual models;
• low-resource NLP research.

It is also one of the token-classification tasks included in the SINHALA-GLUE benchmark.

Source Data

The sentences are drawn from Sinhala news text.

The original dataset was collected and annotated by Manamini et al. (2016) for developing and evaluating the Ananya Sinhala named entity recognition system.

This repository does not introduce a new annotation effort. It is a re-upload of the original dataset, prepared for easier access and its use within SINHALA-GLUE.

Languages

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

Limitations

The dataset is based on Sinhala news text and therefore primarily represents the vocabulary, entities and writing styles found in the news domain.

The annotation scheme contains three named entity categories: persons, organisations and locations. Other entity types commonly included in broader NER schemes are not represented.

As an older dataset collected from a particular domain, its entity distribution may not reflect contemporary Sinhala usage or named entities occurring in other domains such as social media, conversational text, legal documents or biomedical text.

Users should therefore consider the domain and annotation coverage when interpreting results obtained using this dataset.

Citation

As this repository is a re-upload of an existing dataset, please cite the original paper that introduced the dataset:

@inproceedings{manamini2016ananya,
    title = {Ananya - a Named-Entity-Recognition (NER) System for Sinhala Language},
    author = {Manamini, S. A. P. M. and
      Ahamed, A. F. and
      Rajapakshe, R. A. E. C. and
      Reemal, G. H. A. and
      Jayasena, S. and
      Dias, G. V. and
      Ranathunga, S.},
    booktitle = {2016 Moratuwa Engineering Research Conference (MERCon)},
    year = {2016},
    pages = {30--35},
    publisher = {IEEE},
    doi = {10.1109/MERCon.2016.7480111}
}

The original paper was published at the 2016 Moratuwa Engineering Research Conference (MERCon), pages 30–35.

If you use the SINHALA-GLUE version or splits provided in this repository, please also cite:

@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 included as the Named Entity Recognition (NER) task in SINHALA-GLUE, a Sinhala natural language understanding benchmark.

SINHALA-GLUE includes:

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