UD_Thai-TUD

Treebank card from UD release 2.18: metadata and texts come from UD_Thai-TUD/README.md; size is counted by streaming the .conllu files and checked against the release stats.xml. Generated by tools/treebanks build; do not edit by hand. Papers: Papers: UD_Thai-TUD; seeds: seeds/; converter: convert.md.

Annotation layers

UD scale: manual native — annotated by hand directly in UD; converted from manual — converted from manual annotation in another scheme; converted with corrections — converted and partly corrected by hand; automatic with corrections — automatic, with partial manual corrections; automatic — automatic; not available — the layer is absent.

Size

Tokens are surface tokens (Spanish vámonos is one), words are syntactic words (vamos + nos); multiword tokens are those split into several words; empty nodes are the N.M nodes of the enhanced graph.

Summary

UD Thai TUD (Thai Universal Dependency Treebank) is a treebank of 3,627 syntactic trees from the Thai National Corpus and Wikipedia, annotated in Universal Dependencies, covering diverse text types and topics across various domains.

Introduction

The UD Thai TUD treebank was created to provide a broad-coverage syntactic resource for the Thai language under the Universal Dependencies (UD) framework. Text was randomly sampled from two major sources: the Thai National Corpus and the November 2020 dump of Thai Wikipedia. To ensure diversity, 5,000 paragraphs were selected from various document types—news articles, Wikipedia entries, essays, advertisements, interviews, and stories—covering a wide range of topics such as politics, crime, entertainment, sports, history, religion, culture, and science. After annotation and rigorous quality control, 3,627 well-formed dependency trees were retained in the final dataset.

Annotation Process

All paragraphs were tokenized using the newmm tokenizer from the PyThaiNLP library, then annotated on the Datasaur platform. A team of 10 annotators with linguistics backgrounds was trained to: (1) correct tokenization errors, (2) assign Universal POS (UPOS) tags, (3) identify dependency arcs, and (4) label relations (DEPREL) without subtypes. The LEMMA field was excluded due to the lack of inflectional morphology in Thai.

Following pilot annotations and manual review, two annotators demonstrating the highest accuracy were selected to complete the remaining data. Agreement was evaluated on 20 double-annotated sentences (399 tokens), achieving Cohen’s Kappa scores of 0.92 (UPOS) and 0.84 (DEPREL), and UAS/LAS scores of 0.85 and 0.78. The annotated data was then converted into CoNLL-U format and split into individual trees based on dependency structure. Trees with incomplete labels, multiple roots, or structural errors were corrected, with additional quality assurance performed via manual inspection of 50 randomly selected trees.

The treebank consists of randomly shuffled sentences sampled from the Thai National Corpus (TNC) and the November 2020 dump of Thai Wikipedia, rather than complete documents. Each filename encodes the source document and the portion of the document from which the sentences were extracted:

• Wikipedia trees: filenames follow the format wiki_<wgArticleID>.
• TNC trees: filenames follow the format [tnc/][Original TNC filename][Part of the document].

Train–Test Split

The final treebank was split into training, development, and test sets in an 8:1:1 ratio. It consists of syntactically complete trees rather than full documents. While Filenames are used to identify source paragraphs, typically reflecting their origin from the Thai National Corpus or Wikipedia. However, sentence IDs in the final treebank do not encode genre or domain metadata.

<center>

UPOS	Train	Dev	Test	UPOS	Train	Dev	Test
NOUN	18777	2270	2310	CCONJ	2063	239	270
VERB	14881	1802	1867	ADJ	1575	223	197
ADP	4517	530	560	PART	1366	156	169
ADV	4498	557	521	NUM	1161	165	118
AUX	3424	401	421	DET	1140	137	144
PRON	2796	322	350	PUNCT	871	104	125
SCONJ	2438	321	335	SYM	16	1	1
PROPN	2488	293	295

Table 1. UPOS Distribution in Each Split of TUD

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/treebanks/th/tud/