ThaiTrees: Thai Syntactic Dependency Trees Across Domains

• Manual syntactic annotation is costly to create and maintain at scale, while parsed corpora make grammatical relations queryable for distributional and alternation analyses.Source: §1 · p. 1
• Dependency grammar represents syntax as labeled head-dependent relations between words, providing a convenient basis for computational analysis.Source: §1 · p. 1
• ThaiTrees provides 342M tokens from four domains, automatically parsed under Universal Dependencies, with a reproducible pipeline and machine-readable lexicon and CoNLL-U parses.Source: §1 · p. 1

• Thai-TUD provides 3,627 manually annotated trees for parser training and evaluation, but existing Thai resources are limited in scale, register coverage, or dependency annotation.Source: §2 · pp. 1–2
• Large parsed corpora support corpus-linguistic research elsewhere, but Thai lacks a comparably broad resource with queryable dependency analyses.Source: §2 · p. 2
• Thai annotation tools include AttaCut, CRFcut, AttaParse, and a PhayaThaiBERT-based UPOS tagger, with parser models trained on Thai-TUD (Chormai et al., 2020; Phatthiyaphaibun et al., 2023; Sriwirote et al., 2025, 2024).Source: §2 · p. 2

3.1 Data Sources

• ThaiTrees draws on journalistic news, Wikipedia, conversational spoken transcripts, and informal social-media writing; spoken documents are fewer but roughly an order of magnitude longer.Spoken sources include manually transcribed podcasts, film subtitles, and livestreams; social media combines sentiment-dataset and forum posts.Source: §3.1 · p. 2

3.2 Processing Pipeline

• The lexicon and parsing branches segment different text units—whole documents versus individual sentences—so their token boundaries and token identifiers can differ.Source: §3.2 · p. 2, Figure 1 · p. 3

3.2.1 Word segmentation

• AttaCut segments whole documents for the lexicon branch, which filters punctuation, numeric tokens, emoji, whitespace-only tokens, and forms occurring at most five times.Source: §3.2.1 · p. 3

3.2.2 Sentence segmentation and tokenization

• The parsing branch normalizes whitespace, caps social-media documents at 500,000 characters, and uses CRFcut to return sentence strings without carrying its internal token boundaries forward.The cap truncates 149 long, concatenated documents.Source: §3.2.2 · p. 3

3.2.3 Dependency parsing

• AttaParse receives AttaCut tokens as pretokenized sentences and predicts labeled dependency arcs while leaving POS and lemma processing disabled.Source: §3.2.3 · p. 3

3.2.4 POS tagging and benchmark performance

• 90.64% held-out accuracy and 81.34% macro F1 across 15 UPOS classes are reported for the PhayaThaiBERT tagger, whose predictions replace the UPOS column without changing dependency edges.The best checkpoint was selected by held-out accuracy.Source: §3.2.4 · p. 4, Table 2 · p. 4

3.3 Released Artifacts

• The releases include raw text, a frequency lexicon, CoNLL-U parses, and per-pattern edge counts, linked by document IDs but not token IDs.The lexicon contains 451,252 word forms; the parsed corpus contains 203,892,200 tokens and 199,836,464 dependency edges.Source: §3.3 · p. 4

4.1 Methodology

• The log odds ratio ranks words by how much their frequency differs between a target domain and the union of the other three, rather than by frequency alone.The paper reports its natural logarithm to make the measure symmetric around zero; positive values indicate concentration in the target and negative values under-representation.Source: §4.1 · pp. 4–5
• Keyness scoring requires at least 10,000 occurrences in both target and reference and appearances in more than ten target-domain documents.The scoring also excludes single-character forms, ASCII-punctuated tokens, URL fragments, non-Thai/non-Latin-only tokens, and pure-Latin tokens shorter than three characters.Source: §4.1 · p. 5

4.2 Results and Discussion

• The corpus-wide frequency distribution is Zipfian, and its top 20 words are predominantly function words and other forms with broad grammatical or combinatory roles.Examples include verbal forms with ability, deictic, copular, or preposition-like uses, and productive nominal forms; the same high-frequency forms recur across domains with slightly different rankings.Source: Figure 2 · p. 5, §4.2 · p. 6
• News keyness highlights law, government, and incident-reporting vocabulary, including police, case, legal-section, announce, and confirm.The keyness lists compare each domain against the other three combined and apply frequency and document-count thresholds.Source: §4.2 · pp. 6–7, Table 5 · p. 6
• Wikipedia favors dates, names, and sport or entertainment terms; spoken language favors conversational particles and informal pronouns; social media features finance, commerce, and buyer–seller politeness.The reported examples include football and album for Wikipedia, and stocks, bank, buy, and thank you for social media.Source: §4.2 · p. 7

5.1 Methodology

• Each counted pattern is a dependency edge represented by the head POS, Universal Dependencies relation, and dependent POS, rather than a subtree.The methodology applies keyness calculations to syntactic patterns and retains patterns with at least 10,000 edges in both target and reference domains, with no minimum document count.Source: §5.1 · pp. 7–8, Table 7 · p. 8

5.2 Results and Discussion

• Figure 3 illustrates the five most frequent patterns, including noun modification, verb–object relations, and compounding.The illustrated word pairs are corpus instances selected from the most frequent pairs for each pattern.Source: §5.2 · p. 8
• The 20 most frequent patterns are familiar Thai constructions despite parser attachment accuracy below 90%, supporting aggregate downstream analyses.Twelve patterns are verb-headed and eight noun-headed; NOUN-nmod-NOUN and VERB-obj-NOUN lead their respective groups.Source: §5.2 · p. 8
• 9.7M instances make third-ranked VERB-compound-VERB a common serial-verb construction; fourth-ranked NOUN-acl-VERB covers verb modification, including relative and nominalized clauses.Given the frequency of การ and ความ, many NOUN-acl-VERB instances likely involve nominalization rather than relative clauses.Source: §5.2 · p. 8
• The remaining frequent patterns include adverbial modification, oblique nominal dependents, auxiliaries, and subordinate clauses.Together, the top-20 patterns present common Thai grammatical constructions despite imperfect automatic parsing.Source: §5.2 · p. 8
• News and Wikipedia cluster together, while spoken transcripts and social media form another group in the top-20 domain rankings.Spoken rankings place VERB-advmod-ADV, VERB-nsubj-PRON, and VERB-aux-AUX higher; VERB-advmod-ADV often involves ก็, which can signal discourse relations.Source: §5.2 · pp. 8–9

6 Conclusion

• ThaiTrees contains 342M tokens across four domains to measure variation in word and syntactic frequency.The corpus is available in machine-readable formats with a frequency lexicon.Source: §6 · p. 9
• News and Wikipedia form a formal pair, while spoken language and social media form a conversational pair in both word and syntactic frequency.Source: §6 · p. 9