sukity/thai-sentence-boundaries

Thai Sentence Boundaries — a small permissive gold set

Licensing note (repo label vs. per-record). This card is labeled cc-by-2.0 as the single most-binding license across the set so downstream use is always safe; most records are actually CC0 (see the license field on each row and the License & attribution section). Only the tatoeba_synth split carries CC-BY (attribute Tatoeba); everything else is public-domain / CC0.

A small, fully permissive gold set of Thai sentence boundaries — 905 hand-verified sentences across 8 registers, released so anyone can evaluate Thai sentence segmenters without a licensing tangle. Every record is CC0 or CC-BY (Tatoeba); there are no non-commercial, no-derivatives, share-alike, or gated parts.

Why. The established Thai sentence-boundary corpora — ORCHID (CC-BY-NC-SA), LST20 / Blackboard Treebank (NECTEC, gated, redistribution restricted) — are non-commercial or gated. We wanted something small but fully permissive (CC0 / CC-BY) to evaluate on and share freely, so we made and released this. It's a starting point, not a comprehensive corpus — use it, fork it, add to it.

---

---

What's in it

register	docs	sentences	license	one-line
handgold_prose	16	182	CC0-1.0	hand-annotated free prose (public-domain literature + PD statutes)
tatoeba_synth	150	592	CC-BY-2.0-FR	native Thai Tatoeba sentences concatenated → exact boundaries
formal_news	1	22	CC0-1.0	formal news register
academic_legal	1	21	CC0-1.0	academic / legal-statute register
novel_dialogue	1	23	CC0-1.0	narrative + nested dialogue quotes
social_chat	1	22	CC0-1.0	informal chat, no terminal punctuation
poetry_lyrics	5	21	CC0-1.0	poems / lyrics with line breaks
tricky_tokens	1	22	CC0-1.0	abbreviation / date / decimal / URL false-boundary traps
total	176	905		26 docs CC0 · 150 docs CC-BY

Each register lives in data/<register>.jsonl; data/all.jsonl concatenates all of them (the register field lets you split back out).

The registers, in detail

Thai has no obligatory sentence-terminating punctuation and uses spaces both between words/phrases and between sentences, so sentence segmentation is a real disambiguation task. These registers were chosen to stress the parts that make it hard, from different angles:

• handgold_prose (182) — the flagship set: real, connected prose, not synthetic. Two sub-sources, both public domain:
  • Classical literature: นิทานเวตาล (translated by Prince Bidyalongkorn, d. 1945) and ราชาธิราช (translated by Chaophraya Phrakhlang (Hon), d. 1805) — out of copyright by author-death term. Narrative prose with embedded dialogue.
  • Statutes: the preamble of the Civil and Commercial Code (B.E. 2466 / 1923) and the Interim Constitution (B.E. 2557 / 2014) — statutes, regulations and gazette text are not subject to copyright under the Thai Copyright Act B.E. 2537, Section 7. Long, clause-dense formal sentences.
• tatoeba_synth (592) — each document is several independent, native Thai Tatoeba sentences joined with single spaces. Because we control the join, the boundaries are exact (100% correct). This gives clean, high-volume boundary signal and, crucially, natural intra-sentence spaces (e.g. a space before ๆ) that must not be cut — a good false-positive test.
• formal_news (22) — long noun-phrase sentences with few lexical end cues; where "cut every space" over-segments and models must learn restraint.
• academic_legal (21) — มาตรา / พ.ศ. / ordinal section numbers, whose periods are candidate false boundaries.
• novel_dialogue (23) — narration interleaved with dialogue in nested quotation marks (“ ” « » 「 」), including a quote-inside-a-quote case.
• social_chat (22) — colloquial, run-on, no terminal punctuation — boundaries carried only by particles/context.
• poetry_lyrics (21, 5 poems) — line breaks that are sometimes sentence ends and sometimes enjambment; deliberately mixes both.
• tricky_tokens (22) — deliberate traps that must not trigger a boundary: ด.ช. / พ.ศ. / ค.ศ. / ม.ค. / น.ส. / กทม. / ดร., decimals (3.14), dates (5 ม.ค. 2567), URLs, and mixed Thai-English-Chinese.

How it was made (provenance & method)

• Hand-written sets (formal_news, academic_legal, novel_dialogue, social_chat, poetry_lyrics, tricky_tokens): authored by hand for evaluation, each targeting the phenomenon described above. Original work, released CC0.
• handgold_prose: sentences were split by reading the public-domain source and then reconstruction-verified programmatically against the exact source string (byte-for-byte) to catch transcription slips. The annotation rulebook covers: dialogue with/without quotation marks; question/imperative-final particles; enumerated clauses kept as one sentence vs. split; multi-token dates and eras; and ฯลฯ (et cetera — sentence-internal) vs. a mid-clause honorific ฯ such as โปรดเกล้าฯ / กรุงเทพฯ (never a boundary). One passage with source OCR corruption was excluded rather than silently fixed.
• tatoeba_synth: built by concatenating distinct native Thai Tatoeba sentences with single spaces. The underlying sentences are held out from any model training we did, so it is safe as an evaluation split. Boundaries are exact by construction.
• Validation: every record's boundaries are strictly ascending and the final boundary equals the document's rune length; the provided loader reconstructs all 176 documents without error.

We deliberately used no NC / ND / SA or gated corpora (no ORCHID, LST20, TED, FLORES) at any point, so this set is free to redistribute and build on.

Data format

One JSON object per line:

{
  "text": "<full document text, exactly as it should be segmented>",
  "sentences": ["<sentence 1>", "<sentence 2>", "..."],
  "register": "handgold_prose",
  "license": "CC0-1.0",
  "source": "hand-annotated PD/CC0 prose"
}

sentences are the ordered gold sentences that make up text. We ship the raw text and the sentences list (rather than pre-computed integer offsets) because some registers contain intra-document line breaks and non-space joins (poetry, dialogue); this way you derive boundary positions under your own tokenization and metric, with no lossy normalization on our side. A gold boundary sits at the end of each sentence.

How to use it

Derive boundary positions (rune indices) from text + sentences:

import json

def gold_cuts(rec):
    """Rune index after which each gold sentence ends (last == len(runes))."""
    text, cuts, pos = list(rec["text"]), [], 0
    for s in rec["sentences"]:
        # skip any leading whitespace/newline, then consume the sentence
        while pos < len(text) and "".join(text[pos:pos+len(s)]) != s and text[pos].isspace():
            pos += 1
        pos += len(s)
        cuts.append(pos)
    return cuts

for line in open("data/all.jsonl", encoding="utf-8"):
    rec = json.loads(line)
    text, gold = rec["text"], set(gold_cuts(rec))
    # compare your segmenter's predicted cut positions against `gold`

Metrics. The two standard measures for Thai sentence segmentation:

• Space-correct accuracy — accuracy over inter-word-space decision points (did we correctly keep vs. cut each space?). Dominated by non-boundaries, so it rewards precision on the majority-negative spaces.
• Boundary-class F1 — precision/recall/F1 over sentence-end positions. More sensitive to recall on the (minority) true boundaries.

Report both — they can disagree, and a segmenter tuned for one may trade the other.

Reference results

Measured with thai-nlp-go's sentence/charseg (character-level, pure-Go, no tokenizer) and sentence/crf (a port of PyThaiNLP's crfcut) engines. Macro-averaged boundary-F1, equal weight per register, over the permissive sets here:

engine	macro boundary-F1	note
whitespace (cut every space)	0.666	recall 1.0, low precision
crf.Default (crfcut port)	0.559	word-level CRF, needs a tokenizer
charseg.Default	0.663	char-level, no tokenizer, ~24× faster than the CRF
charseg, recall-lean τ	0.698	operating-point tuned on held-out data

These are honest and register-dependent — no single engine wins everywhere. Formal news favors "cut every space"; dialogue, tricky-tokens and the free-prose gold favor the learned model. On the out-of-domain academic ORCHID corpus (which this set does not include, being NC), a same-data head-to-head has the char-level model within ~0.01–0.02 boundary-F1 of the word-CRF at ~24× the speed. Treat the numbers above as a starting reference, not a leaderboard.

Honest limitations

• Small. 905 sentences total — a seed/eval set, not a training corpus or a comprehensive benchmark. ORCHID (30k sentences) and LST20 (74k) are far larger (but NC/gated).
• Register imbalance. The six hand-written registers are ~20 sentences each (one document apiece, except poetry with five); tatoeba_synth and handgold_prose are larger. Macro-averaging treats each register equally to compensate, but small per-register counts mean per-register numbers are noisy.
• tatoeba_synth is synthetic — real sentences, but the document boundaries come from our concatenation, not from naturally occurring multi-sentence text.
• Single-source annotation. Boundaries reflect one annotation pass with the documented rulebook; there is no multi-annotator agreement study. Sentence segmentation has genuine gray areas (clause vs. sentence, enumerations) — we aimed for consistency, not a claim of ground truth.

License & attribution

Per-record, via the license field. No NC / ND / SA / gated components.

• CC0-1.0 — all hand-written sets and handgold_prose. Public-domain dedication: copy, modify, distribute, use (even commercially), no attribution required. Underlying handgold_prose texts are themselves public domain (classical literature by age; Thai statutes are non-copyright under the Thai Copyright Act B.E. 2537 §7); the annotation is our original CC0 work. https://creativecommons.org/publicdomain/zero/1.0/
• CC-BY-2.0-FR — the tatoeba_synth split. The sentence text is from Tatoeba (CC-BY 2.0 FR): attribute Tatoeba if you use it. The concatenation + boundary annotation are ours (CC0), but the underlying sentences remain Tatoeba's under CC-BY. https://creativecommons.org/licenses/by/2.0/fr/

Filter on the license field if you need a strictly CC0-only subset.

Contributing

More permissive registers are very welcome — especially larger natural free-prose from public-domain or CC0/CC-BY sources, or additional annotators on the existing docs. Open an issue or PR. Please keep everything CC0 / CC-BY / PD; no NC / ND / SA / gated material.

Citation

Thai Sentence Boundaries — a small permissive gold set.
https://github.com/capytron/thai-sentence-boundaries
Companion to the thai-nlp-go project: https://github.com/sukitss/thai-nlp-go