Nevidu/SiDiaC-v.2.5

Sinhala Diachronic POS & Lemma Corpus

Token-level part-of-speech tags and lemmas for a corpus of 59 OCR'd historical Sinhala documents (5th–20th century), built as part of a study on diachronic semantic change in Sinhala (see Paper below). Provided in two formats:

• data/sinhala_diachronic_pos_lemma.conllu — CoNLL-U Plus, the standard format for POS/lemma-annotated corpora in this field (loadable with conllu, spaCy, Stanza, UDPipe, etc.).
• data/sinhala_diachronic_pos_lemma.csv — the same data as a flat CSV, for anyone who'd rather not parse CoNLL-U.

Paper

Dynamics of meaning: Towards the Evaluation of Diachronic Semantic Change in Sinhala Nevidu Jayatilleke, Nisansa de Silva (2026)

📄 arXiv:2609.08609

Loading the dataset

CSV, via 🤗 datasets:

from datasets import load_dataset

ds = load_dataset("Nevidu/SiDiaC-v.2.5")
print(ds)
print(ds["train"][0])

CSV, via pandas (if you just want a DataFrame):

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Nevidu/SiDiaC-v.2.5",
    repo_type="dataset",
    filename="data/sinhala_diachronic_pos_lemma.csv",
)
df = pd.read_csv(path)

CoNLL-U Plus (pip install conllu):

from huggingface_hub import hf_hub_download
from conllu import parse_incr

path = hf_hub_download(
    repo_id="Nevidu/SiDiaC-v.2.5",
    repo_type="dataset",
    filename="data/sinhala_diachronic_pos_lemma.conllu",
)

with open(path, encoding="utf-8") as f:
    for sentence in parse_incr(f):
        print(sentence.metadata["sent_id"], sentence.metadata["text"])
        for token in sentence:
            print(token["form"], token["lemma"], token["upos"], token["xpos"])
        break  # remove to iterate the whole corpus

Dataset structure

One row/token line per token, grouped into sentences. Columns (CSV) / per-token fields (CoNLL-U, via # global.columns):

column	description
century	Century the source document was written in (5, 12, 13, 14, 15, 16, 18, 19, 20 — no attested documents in the 17th century)
sentence_index	Index of the sentence within its century's document collection
token_index	Position of the token within its sentence (0-indexed)
word	Surface form (FORM)
lemma	Finalized lemma (see "How POS/lemma were finalized" below)
pos	Finalized POS tag, SinMorphy's own category labels (this dataset's XPOS)
upos	pos mapped to the Universal POS tagset
sinmorphy_pos, sinmorphy_lemma, sinmorphy_morphology	SinMorphy's own single-best pos/lemma/morphology tag (e.g. කාම+N+NUTR+PL+NOM) before ensemble resolution
sinmorphy_all_tags, sinmorphy_all_morphologies	SinMorphy's full candidate list when ambiguous (|-separated)
sinling_pos, tnt_pos	The other two taggers' individual votes
doc_id, doc_title, doc_title_en, doc_author, doc_genre	Source document metadata
written_date_start, written_date_end, issued_date, ocr_confidence	Dating and OCR quality metadata

CoNLL-U sentence blocks additionally carry # sent_id, # text, # century, # doc_id, # doc_title, # doc_author, # written_date comments. FEATS/HEAD/DEPREL/DEPS/MISC are unspecified (_) — no morphological feature mapping or dependency parse is available in this pipeline.

Row counts per century:

century	rows
5	565
12	866
13	13,617
14	5,210
15	8,483
16	1,369
18	4,707
19	12,701
20	17,295

Total: 64,813 tokens, 4,394 sentences, 59 documents.

upos distribution: NOUN 31,890 · VERB 9,963 · PART 9,329 · X 9,160 · PROPN 2,264 · ADJ 1,661 · PRON 546. doc_genre spans 8 categories (Non-Fiction; Religious, Fiction; Poetry, Non-Fiction; Language, Fiction; Religious, Non-Fiction; History, Non-Fiction; Poetry, Non-Fiction; Medical, Non-Fiction).

How pos and lemma were finalized

Three independent taggers were run over the same tokenized corpus:

1. SinMorphy — a Sinhala morphological analyzer (web tool), scraped via Selenium (Morphological Analysis/sinmorphy_automation_v6.ipynb). Returns one or more candidate (pos, morphology) pairs per word.
2. sinling — a Sinhala NLP library's POS tagger (POS_Tagging/sinling_pos_tagger.ipynb).
3. TnT — a beam-search variant of the Trigrams'n'Tags statistical POS tagger, using a pretrained Sinhala model (POS_Tagging/TnT_Tagger_Robust_Alignment.ipynb).

For words where SinMorphy returned multiple candidates, the candidate is chosen by priority vote: prefer the one matching TnT's tag, else the one matching sinling's tag, else the first candidate. lemma is then the stem extracted from the chosen candidate's morphology string (text before the first +). All three taggers' individual outputs are kept in the dataset (see columns above) so this resolution can be audited or redone.

upos is a fixed, non-destructive mapping from the finalized pos: NOUN→NOUN, PROPER NOUN→PROPN, Verb→VERB, PART→PART, ADJECTIVE→ADJ, PRONOUN→PRON, Proper noun or Invalid→X, FOREIGN WORD→X.

Reproducible end-to-end in build_dataset.ipynb.

How document metadata was attached

Data/century_tokens.json (upstream) concatenates each document's sentences into a per-century list without preserving document boundaries. Document provenance is recovered by re-tokenizing each source document (same function as Data Processing/Data_Partitioning.ipynb) and using each document's own sentence count as a contiguous offset range within its century — documents are concatenated in a fixed traversal order and each contributes exactly one uninterrupted block, so content doesn't need to match exactly for this to be correct, only the sentence counts per document.

Caveat: the current .txt files in Data/OCR_Final/ do not byte-for-byte reproduce century_tokens.json for every document — a small number of sentences differ by a token (both files entered the repo in the same commit and neither has been modified since, so this is a pre-existing snapshot mismatch, not introduced by this notebook). This doesn't affect doc_id/provenance assignment (which relies only on sentence counts, not content), but it means a handful of word/lemma values in this dataset may differ by a token or two from what you'd get re-tokenizing today's OCR_Final/ files directly. Of 59 documents, 4 show a first/last-sentence mismatch against the frozen corpus (flagged in build_dataset.ipynb as "anchor drift" — informational only; total sentence-count accounting for every century balances exactly, so document attribution itself is unaffected).

Known limitations

• ~14.1% of tokens (9,143 rows) have no SinMorphy analysis at all (unrecognized/OOV words, mostly proper nouns) — pos is Proper noun or Invalid (upos = X) and lemma falls back to the raw word.
• 0.3% of tokens (203 rows) hit a SinMorphy morphology-string edge case with no lemma stem before the first + — lemma also falls back to the raw word for these.
• POS/lemma quality depends on the accuracy of the three underlying taggers, which have not been independently evaluated against a gold-standard Sinhala treebank as part of this pipeline.
• FEATS (case, number, tense, etc.) is not populated — SinMorphy's raw morphology tag is preserved as-is in sinmorphy_morphology/SinMorphyMorphology instead of being mapped to UD's Feature=Value syntax.
• License is unspecified — set the license field above (and this notice) before publishing to the Hub.

Provenance

Source texts: Data/OCR_Final/ (documents with per-document metadata: title, author, genre, written/issued date, OCR confidence).

Citation

@article{jayatilleke2026dynamicsmeaningevaluationdiachronic,
      title={Dynamics of meaning: Towards the Evaluation of Diachronic Semantic Change in Sinhala}, 
      author={Nevidu Jayatilleke and Nisansa de Silva},
      year={2026},
      eprint={2609.08609},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.08609}, 
}