NethranjaliSE/pali-sinhala-parallel-18k

Pali–Sinhala Parallel Corpus (Tipiṭaka)

The first publicly available parallel corpus for Pali-to-Sinhala machine translation, comprising 18,455 aligned sentence pairs drawn from three Nikāyas of the Sutta Piṭaka. Pali is written throughout in Sinhala script, as is conventional in the Sri Lankan Theravāda tradition.

---

Dataset Description

Summary

Property	Value
Language pair	Pali (pi) → Sinhala (si)
Pali script	Sinhala (not Roman transliteration)
Total sentence pairs	18,455
Source	pitaka.lk — Digital Tipiṭaka
Domain	Canonical Buddhist prose (Sutta Piṭaka)
License	CC BY 4.0

Corpus Composition

Source	Sentence pairs
Aṅguttara Nikāya	5,604
Khuddaka Nikāya	8,065
Majjhima Nikāya	4,786
Total	18,455

Lexical Statistics

Statistic	Pali (Sinhala script)	Sinhala
Vocabulary size	13,153	11,767
Total running words	76,982	116,426
Type–token ratio	17.09 %	10.11 %
Mean sentence length	4.17 words	6.31 words

The asymmetry is informative. Pali conveys comparable content in roughly two-thirds of the running words that Sinhala uses, while drawing on a larger vocabulary — a consequence of compounding and sandhi that produce a high proportion of rare inflected forms. This is the empirical basis for applying subword segmentation rather than word-level tokenisation.

Data Splits

The corpus is provided with a pre-defined 80 / 10 / 10 split generated under a fixed random seed, so results can be reproduced.

Split	Pairs
train	14,764
validation	1,845
test	1,846

---

Dataset Creation

Source Material

Text was extracted from the digital Tipiṭaka published at pitaka.lk, which presents the Pali canon together with its established Sinhala translation. This edition was selected over alternatives because it applies consistent Unicode encoding throughout; other available digital editions were found to contain legacy non-Unicode font mappings and irregular zero-width joiner usage that produce distinct byte sequences for visually identical characters.

Alignment Procedure

The canonical numbering of discourses and sections is shared between the Pali and Sinhala presentations, providing sentence-level anchor points that ordinary comparable corpora lack. Candidate pairs were extracted at these boundaries and aligned at sentence level. Pairs that could not be confidently aligned were discarded rather than retained with uncertain correspondence.

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("NethranjaliSE/pali-sinhala-parallel-18k")

Associated Models and Results

Models trained on this corpus are available on the Hugging Face Hub.

Model	BLEU	chrF	chrF++
LSTM encoder–decoder	8.10	24.29	—
GRU encoder–decoder	13.99	28.87	—
Transformer (from scratch)	22.96	46.44	—
mBART-50 fine-tuned	26.28	51.34	48.06

All scores are reported on the held-out test split using sacreBLEU with recorded metric signatures.

• Fine-tuned model: NethranjaliSE/Pali_Sinhala_fine_tune_Model
• Interactive demo: NethranjaliSE/Pali_Sinhala_Fine_Tune_Translator

---

Limitations

Domain. The corpus is drawn entirely from the Sutta Piṭaka. Performance on Abhidhamma, Vinaya, commentarial or non-canonical Pali cannot be inferred from these results.

Sentence-level partitioning. The corpus was split at sentence level rather than by complete discourse. Canonical Pali contains extensive formulaic repetition, so near-identical passages may appear in more than one split. Duplicate pairs were removed before splitting; near-identical formulaic variants may still straddle the boundary.

Short sentences. The mean Pali sentence length is 4.17 words. BLEU's geometric mean over n-gram orders produces near-zero scores for adequate translations of very short sentences, which make up approximately half the test set. chrF and chrF++ are more informative metrics for this corpus.

Single direction. Only Pali → Sinhala translation has been evaluated.

Script convention. The Pali text uses Sinhala script throughout. Roman transliteration and Burmese-script editions of the same texts are not covered.

---

Citation

If you use this dataset, please cite:

@misc{Nethranjali2026palisinhala,
  title        = {Neural Machine Translation Approach for Pali to Sinhala Translation},
  author       = {Sachini P. Nethranjali},
  year         = {2026},
  institution  = {Department of Statistics and Computer Science,
                  University of Kelaniya, Sri Lanka},
  note         = {Bachelor of Science Honours thesis.
                  Dataset and model available at
                  \url{https://huggingface.co/NethranjaliSE}}
}

---

Dataset Card Authors

P.S.P. Nethranjali

More Information

Email me : nethranjalise@gmail.com