SinhaLegal

SINHALEGAL is a curated corpus of Sinhala legal documents developed to support research in low-resource Natural Language Processing (NLP), with a particular focus on legal text processing, summarization, and language modeling. The dataset consists of officially published Sri Lankan legal documents that were processed using Optical Character Recognition (OCR) using Google Document AI, and extensively post-processed to ensure textual accuracy and consistency.

GitHub Repository: https://github.com/Minduli-Lasandi/SinhaLegal

---

Dataset Composition

The SINHALEGAL corpus contains a total of 1,206 legal documents, distributed as follows:

• Acts: 1,065 documents
• Bills: 141 documents

All documents are in Sinhala and originate from official legal publications.

---

Dataset Organisation

The dataset is organised hierarchically to enable systematic access and reproducibility.

Directory Structure

• Documents are first categorised by document type:

  • Acts/
  • Bills/

• Within each document type, documents are organised into year-wise folders based on the year of publication.

• Each legal document is stored in a separate directory named after the document and contains:

  • The full cleaned text of the legal document
  • A corresponding metadata file containing structured descriptive information

---

Metadata Description

Each document is accompanied by a metadata file describing the document and its OCR process.

Metadata Fields

The metadata files include:

• Document ID
• File name
• Document type (Act / Bill)
• Year of publication
• Total number of pages
• Number of OCR chunks
  • One chunk corresponds to a maximum of 15 pages
• Overall OCR confidence score

---

Document Post-Processing

The following corrections were applied:

• Word-level corrections
  Correction of misspelt words, broken words, and incorrect substitutions caused by OCR errors.

• Removal of footer content and page numbers
  Footers and page numbers that disrupted paragraph flow were removed.

• Removal of extraneous sentences
  Small, irrelevant sentences appearing outside paragraph boundaries were removed to maintain textual coherence and accuracy.

• Removal of seal content and price information
  Official stamps, watermarks, and seal artifacts were removed. Prices appearing at the beginning of documents were also removed.

• Removal of repeated titles
  Titles appearing multiple times per page were removed, while titles on the title page and the first page were retained.

• Correction of spacing errors
  Inconsistent spacing between sentences and paragraphs was corrected.

• Removal of unnecessary characters
  Extraneous characters such as underscores and dashes introduced during OCR were removed where they were not related to document content.

These post-processing steps were applied to all 1,065 Acts and 141 Bills in the SINHALEGAL corpus.

---

Usage

You can load the dataset using the following code:

from datasets import load_dataset

ds = load_dataset("Minduli-Lasandi/SinhaLegal", split="train")

# Verify the number of rows
print(f"Number of rows: {len(ds)}")

# Print the first sample
print(ds[0])

---

Citation

@misc{lasandi2026sinhalegalbenchmarkcorpusinformation,
      title={SinhaLegal: A Benchmark Corpus for Information Extraction and Analysis in Sinhala Legislative Texts}, 
      author={Minduli Lasandi and Nevidu Jayatilleke},
      year={2026},
      eprint={2603.04854},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.04854}, 
}