📄 The Thai Discourse Treebank: Annotating and Classifying Thai Discourse Connectives

Summary

Thai has had no annotated resource for discourse analysis. The task is also harder in Thai than in English. The writing system does not mark clause or sentence boundaries with punctuation, and the language has little morphology to flag connectives. Many connective words are also ambiguous between senses or can play non-discourse roles. To fill this gap, the authors built the Thai Discourse Treebank, the first sizeable Thai corpus annotated in the Penn Discourse Treebank (PDTB) tradition. They wrote Thai-specific guidelines based on the PDTB-3 scheme and labelled explicit connectives, their senses and paired connectives on tokenized text from the LST20 corpus. Part of the data also has argument spans. They then benchmarked connective identification and sense classification with several systems: a dictionary baseline, CRF, MaxEnt and pretrained language models such as WangchanBERTa. Identification proved nearly solved, but assigning senses remains difficult. Model errors closely mirror the cases where human annotators disagreed. The released corpus and connective inventory offer a baseline for Thai discourse parsing. They may also help with other languages that lack clause marking, such as Lao and Burmese.

Key points

• Corpus size: 384 documents and about 10,600 sentences, with 18,374 connective instances across 33 relation senses. Argument spans are annotated for 180 of the documents.
• Connective inventory: 148 potentially polysemous connectives, giving 340 form–sense pairs, released with classification criteria. The data is available at github.com/nlp-chula/thai-discourse-treebank.
• Sense scheme: the scheme stays close to PDTB-3, adding only Expansion.GenExpansion to keep compatibility.
• Best results: F1 0.96 for connective identification and macro-F1 0.46 for level-2 sense classification. PLMs beat CRF models at every sentence length, and single-sentence input worked better than sentence pairs.
• Error analysis: the main error sources were confusable senses (e.g., Concession.Arg2-As-Denier vs. Contrast), coordination, and words that can serve other grammatical functions. These patterns parallel inter-annotator disagreements.