sl-parliamentary-nlp/sl-parliamentary-hansard-17-26

Dataset Card for Sri Lanka Parliamentary Hansard

Sri Lanka Parliamentary Hansard is a trilingual parliamentary speech corpus built from publicly available Hansard records of the Parliament of Sri Lanka. It contains Sinhala (සිංහල), Tamil (தமிழ்), English, and code-mixed speeches from 2017 to 2026, with speaker names, dates, and topic-modeling labels.

The dataset was created for the research paper "Trilingual Topic Modeling of Sri Lankan Parliamentary Debates", associated with MERCon 2026, and supports work on low-resource multilingual NLP, parliamentary discourse analysis, and unsupervised topic modeling.

This is a research dataset and is not an official account, publication, or endorsement of the Parliament of Sri Lanka.

Dataset Details

Each row represents a speech-level Hansard segment. The topic labels were generated using a multilingual topic-modeling pipeline based on BGE-M3 embeddings, UMAP dimensionality reduction, HDBSCAN micro-topic clustering, and macro-topic aggregation.

Dataset Sources

• Dataset repository: https://huggingface.co/datasets/sl-parliamentary-nlp/sl-parliamentary-hansard-17-26
• Project dashboard: https://hansards.vercel.app/
• Code repository: https://github.com/HimathX/lk-hansard-topic-modeling
• Publication venue: https://mercon.uom.lk/
• Primary source: Publicly available official Hansard PDFs from the Parliament of Sri Lanka archive: https://www.parliament.lk/en/business-of-parliament/hansards

Uses

Direct Use

This dataset is intended for:

• Multilingual NLP research for Sinhala, Tamil, English, and code-mixed text
• Parliamentary discourse analysis
• Topic modeling and clustering research
• Low-resource language benchmarking
• Political science research on agenda setting and temporal attention
• Speaker-level and event-level analysis of parliamentary debate

Out-of-Scope Use

This dataset should not be used as:

• An official parliamentary record
• A source for legal, electoral, or policy decisions without manual verification against the original Hansard documents
• A ground-truth political stance or sentiment dataset
• A complete representation of every parliamentary utterance without accounting for extraction and filtering limitations

Dataset Structure

Column	Type	Description
speech_id	string	Unique speech identifier, such as SP_00000
date	string	Parliamentary session date in YYYY-MM-DD format
speaker	string	Normalized speaker name, which may appear in Sinhala, Tamil, or English
text	string	Full speech text in the original language, including possible code-mixing
year	integer	Year extracted from the session date
micro_topic_cluster	integer	HDBSCAN micro-topic cluster label; -1 indicates procedural noise or an unassigned speech
macro_topic	string	Aggregated macro-topic label, such as Macro-Topic 3, or Procedural Noise

How to Get Started

Install the Hugging Face datasets library:

pip install datasets pandas

Load the dataset with the default configuration:

from datasets import load_dataset

dataset = load_dataset(
    "sl-parliamentary-nlp/sl-parliamentary-hansard-17-26",
    "default",
    download_mode="force_redownload",
)

print(dataset)

Convert to pandas:

train_df = dataset["train"].to_pandas()
test_df = dataset["test"].to_pandas()

df = train_df
print(df[["date", "speaker", "macro_topic", "text"]].head())

Example: count speeches by macro-topic:

topic_counts = (
    df["macro_topic"]
    .fillna("Unassigned")
    .value_counts()
)

print(topic_counts.head(10))

Corpus Construction

The corpus was constructed from official Hansard PDFs available through the Sri Lankan parliamentary archive.

The processing pipeline included:

1. Scraping publicly available Hansard PDFs.
2. Extracting text with Google Gemini, selected because OCR struggles with the dual-column trilingual layout and historical Sinhala/Tamil font encodings.
3. Segmenting extracted text into speech-level units.
4. Normalizing speaker names across Sinhala, Tamil, and English variants.
5. Removing structural headers, footers, and short non-substantive fragments.
6. Assigning topic labels through the multilingual topic-modeling pipeline.

Topic Modeling Method

The topic-modeling pipeline used:

• Embeddings: BGE-M3 dense multilingual embeddings
• Dimensionality reduction: UMAP
• Micro-topic clustering: HDBSCAN
• Macro-topic aggregation: empirical dendrogram cut over micro-topic centroids
• Noise handling: HDBSCAN label -1 for procedural or low-density speeches

The final topic-modeling run produced:

Item	Value
Topic-labeled speeches	19,553
HDBSCAN micro-topics	336
Procedural/noise speeches	6,921
Substantive clustered speeches	12,632
Macro-topic groups	30

Evaluation

The project includes a manually labeled 300-speech evaluation sample and reviewer-facing evaluation artifacts in the companion code repository.

Internal clustering benchmarks reported for the full corpus include:

Algorithm	K	Noise %	Evaluated rows	Silhouette	Calinski-Harabasz	Davies-Bouldin
HDBSCAN	336	35.4%	12,632	0.5741	15,052	0.5148
KMeans	336	0.0%	19,553	0.4076	13,098	0.8745
Agglomerative	336	0.0%	19,553	0.3812	12,217	0.9088

Limitations

• Extraction noise: The source PDFs contain complex formatting, multilingual text, and historical font encodings. LLM-based extraction improves coverage but may still introduce errors.
• Speaker normalization: Names are normalized heuristically, so rare or ambiguous speaker variants may remain unmerged.
• Topic labels are unsupervised: micro_topic_cluster and macro_topic are model-generated analytical labels, not official parliamentary categories.
• Noise labels are real text: Speeches labeled -1 are not invalid records; they are simply not assigned to a dense HDBSCAN topic region.
• Temporal coverage is uneven: Parliamentary sitting frequency differs by year, so raw counts should be normalized before making year-to-year claims.
• Political text can be sensitive: Users should manually verify outputs before using the dataset in high-stakes political, journalistic, legal, or policy contexts.

Bias, Risks, and Considerations

The dataset reflects the language, framing, and political priorities present in parliamentary debate. It may contain partisan language, allegations, sensitive political claims, and socially biased statements made by speakers. Topic models may amplify or obscure these patterns depending on preprocessing choices and clustering parameters.

Researchers should treat the dataset as a computational research resource, not as a neutral summary of Sri Lankan politics.

Citation

If you use this dataset, please cite:

@inproceedings{dhanapala2026trilingual,
  title={Trilingual Topic Modeling of Sri Lankan Parliamentary Debates},
  author={Dhanapala, Himath and Daishika, Haren and Kuruppu, Himandhi and Seneviratne, Sithija and Kavindya, Ashini and Narasinghe, Patalee and Weerasekara, Sandeepa and de Silva, Nisansa and Wickramanayake, Sandareka},
  booktitle={Proceedings of the 12th Moratuwa Engineering Research Conference (MERCon 2026)},
  year={2026},
  publisher={University of Moratuwa, Sri Lanka},
  url={https://mercon.uom.lk/}
}

Dataset Card Authors

This dataset card was prepared by the Sri Lanka Parliamentary NLP project team based on the companion research repository and dataset publishing artifacts.

License

The dataset is released under the Apache 2.0 license. The underlying Hansard text is derived from publicly available parliamentary records of the Parliament of Sri Lanka.