Sequence-Labeling RoBERTa Model for Dependency-Parsing in Classical Chinese and Its Application to Vietnamese and Thai

The author and his colleagues have been developing classical Chinese treebank using Universal Dependencies. We also developed RoBERTa-Classical-Chinese model pre-trained with classical Chinese texts of 1.7 billion characters. In this paper we describe how to finetune sequence-labeling RoBERTa model for dependency-parsing in classical Chinese. We introduce “goeswith”-labeled edges into the directed acyclic graphs of Universal Dependencies in order to resolve the mismatch between the token length of RoBERTa-Classical-Chinese and the word length in classical Chinese. We utilize [MASK]token of RoBERTa model to handle outgoing edges and to produce the adjacency-matrices for the graphs of Universal Dependencies. Our RoBERTa-UDgoeswith model outperforms other dependency-parsers in classical Chinese on LAS / MLAS / BLEX benchmark scores. Then we apply our methods to other isolating languages. For Vietnamese we introduce “goeswith”-labeled edges to separate words into space-separated syllables, and finetune RoBERTa and PhoBERT models. For Thai we try three kinds of tokenizers, character-wise tokenizer, quasi-syllable tokenizer, and SentencePiece, to produce RoBERTa models.