ThaiTrees: Thai Syntactic Dependency Trees Across Domains

Attapol T. Rutherford; Papatchol Thientong

cs.CLSep 23, 2026

ThaiTrees: Thai Syntactic Dependency Trees Across Domains

Authors: Attapol T. Rutherford, Papatchol Thientong

Organizations: Department of Linguistics Chulalongkorn University

Abstract

Studying syntactic patterns in naturally occurring language requires a large parsed corpus, but manual annotation is costly and difficult to scale. Thai has a manually annotated dependency treebank for training and evaluating parsers, but lacks a large automatically parsed corpus for quantitative syntactic research. We present ThaiTrees, a 342M-token corpus drawn from news, Wikipedia, spoken transcripts, and social media. We develop a reproducible pipeline for cleaning, processing, and parsing Thai text under the Universal Dependencies framework. The resulting corpus makes grammatical relations searchable and supports the study of syntactic distributions. We release a frequency lexicon and CoNLL-U parses in machine-readable formats suitable for both AI-assisted and conventional programmatic analysis.

Figures & tables

Domain	Documents	Tokens	Sentences	Source
News	100,846	52,641,127	897,348	ThaiPBS website
Wikipedia	175,069	101,378,583	1,303,443	Thai Wikipedia (MediaWiki random sampling)
Spoken	2,505	30,238,809	266,944	YouTube transcripts (7+ channels)
Social Media	87,700	157,708,614	2,052,040	Wisesight + Pantip datasets
Total	366,120	341,967,133	4,519,775

Table 1: The four sub-corpora that make up ThaiTrees.

Setting	Value
Base model	PhayaThaiBERT
Training data	UD Thai-TUD
Split (sentences)	2,902 / 362 / 363
Split (word tokens)	62,011 / 7,521 / 7,683
Optimizer	AdamW
$\beta_{1},\ \beta_{2},\ \epsilon$	$0.9$ , $0.999$ , $10^{-8}$

Table 2: Fine-tuning configuration for the PhayaThaiBERT POS tagger ( phayathaibert-thai-pos-tagger ). The best checkpoint was selected by held-out accuracy; macro F1 is over the 15 UPOS classes present in the treebank.

Stage	Tool	Benchmark	Source
Word seg.	AttaCut	91 % WL-F1 (BEST)	Chormai et al.,2020
Sent. seg.	CRFcut	87 % / 82 % acc. (ORCHID / TED)	Chumpolsathien,2020
Dep. parsing	AttaParse 1.0	86.4 % UAS / 76.6 % LAS (TUD test) $\dagger$	Sriwirote et al.,2025
POS tagging	PhayaThaiBERT (ours)	90.64 % / 81.34 % macro F1 (UD Thai-TUD)	this work

Table 3: In-domain benchmarks for the four pipeline stages.

\dagger

AttaParse 1.0 = the no-POS graph-based PhayaThaiBERT configuration (row GP) of Sriwirote et al. (2025) , Table 3.

Rank	Count	Freq/M
1	6,300,593	18,424.6
2	3,972,907	11,617.8
3	3,928,739	11,488.6
4	3,729,958	10,907.4
5	3,720,946	10,881.0
6	3,655,798	10,690.5

Table 4: Top 20 words across the corpus by overall frequency per million.

Rank	Word	$\log\mathrm{OR}$	Word	$\log\mathrm{OR}$	Word	$\log\mathrm{OR}$	Word	$\log\mathrm{OR}$
	News		Wikipedia		Spoken		Social Media
1		2.32		2.42		4.07		3.21
2		2.30		2.35		3.28		2.41
3		2.29		2.32		3.25		2.24
4		2.16		2.29		3.11		2.19
5		2.07		2.29		2.74		2.17

Table 5: Top 20 words of each domain by log odds ratio against the union of the other three domains, among words occurring at least 10,000 times in both the target and reference and in more than ten documents of the target domain.

Rank	Pattern	Count	Freq/M
1	NOUN-nmod-NOUN	14,271,548	71,416.1
2	VERB-obj-NOUN	13,929,524	69,704.6
3	VERB-compound-VERB	9,735,805	48,718.9
4	NOUN-acl-VERB	9,083,706	45,455.7
5	NOUN-case-ADP	8,823,314	44,152.7
6	VERB-advmod-ADV	7,575,704	37,909.5

Table 6: Top 20 dependency-edge patterns across the corpus by overall frequency per million. Patterns are written head-relation-dependent .

Rank	Pattern	$\log\mathrm{OR}$	Pattern	$\log\mathrm{OR}$	Pattern	$\log\mathrm{OR}$	Pattern	$\log\mathrm{OR}$
	News		Wikipedia		Spoken		Social Media
1	NOUN-appos-NUM	2.19	PROPN-appos-PROPN	2.93	PART-det-PART	2.80	ADV-punct-PUNCT	1.92
2	NUM-flat-PROPN	1.12	NOUN-appos-PROPN	1.79	PART-amod-PART	2.77	VERB-obl-PUNCT	1.91
3	VERB-compound-CCONJ	0.91	NOUN-nsubj-PROPN	1.65	PART-compound-PART	2.47	ADV-fixed-PART	1.73
4	VERB-cc-SCONJ	0.83	PROPN-conj-NOUN	1.65	CCONJ-fixed-SCONJ	2.23	VERB-discourse-PART	1.72
5	VERB-csubj-VERB	0.75	PROPN-conj-PROPN	1.63	PART-advmod-PART	2.17	ADV-advmod-PART	1.69

Table 7: Top 20 dependency-edge patterns of each domain by log odds ratio against the union of the other three domains. Patterns are written head-relation-dependent (POS of head, dependency relation, POS of dependent); only patterns with at least 10,000 edges in both the target and reference domains are considered.

Explore similar work

CAIT: A Syntactic Parsing Toolkit for Child-Adult InTeractions
May 19, 2026Francesca Padovani, Xiulin Yang, Bastian Bunzeck +4Syntactic Structure Language Acquisition
Center for Language and Cognition (CLCG), University of Groningen · Georgetown University · Computational Linguistics, Department of Linguistics, Bielefeld University
Universal Topological Regularity of Syntactic Structures
Jan 31, 2023Fermín Moscoso del Prado MartínSyntactic Structure Phylogenetic Inference
Department of Language and Communication & Center for Language Studies Radboud University Erasmuslaan 1, 6525 NL, Nijmegen, The Netherlands
AthDGC: An Open Diachronic Greek Treebank with Indo-European Parallels
Jun 13, 2026Nikolaos Lavidas, Kiki Nikiforidou, Dag Haug +8Greek
National and Kapodistrian University of Athens · University of Oslo · Ghent University

May 19, 2026, cs.CLJ/K move · Enter open · S save

CAIT: A Syntactic Parsing Toolkit for Child-Adult InTeractions

Francesca Padovani, Xiulin Yang, Bastian Bunzeck, Jaap Jumelet+3

Center for Language and Cognition (CLCG), University of Groningen · Georgetown University · Computational Linguistics, Department of Linguistics, Bielefeld University

Syntactic Structure Language Acquisition Context-Free Grammars

Open paper AskPDF arXiv

CHILDES is a paramount resource for language acquisition studies -- yet computational tools for analyzing its syntactic structure remain limited. Leveraging the recent release of the UD-English-CHILDES treebank with gold-standard Universal Dependencies (UD) annotations, we train a state-of-the-art dependency parser specifically tailored to CHILDES. The parser more accurately captures syntactic patterns in child-adult interactions, outperforming widely used off-the-shelf English parsers, including SpaCy and Stanza. Alongside the parser, we also release a Part-of-Speech tagger and an utterance-level construction tagger, which together form the open-source Syntactic Parsing Toolkit for Child-Adult InTeractions (CAIT). Through a detailed error analysis and a case study tracking the distribution of syntactic constructions across developmental time in CHILDES, we demonstrate the practical utility of the toolkit for large-scale, reproducible research on language acquisition.

Abstract

Explore similar work

CAIT: A Syntactic Parsing Toolkit for Child-Adult InTeractions

Universal Topological Regularity of Syntactic Structures

AthDGC: An Open Diachronic Greek Treebank with Indo-European Parallels