Extraction and categorization of Japanese lexical collocations with graph-aware transformers

Abstract

Lexical collocations may be identified and categorized in context, which is helpful for language acquisition, dictionary creation, and many other downstream NLP tasks. However, the automatic collocation extraction and categorization using modern machine learning techniques has not been tried in Japanese. In this paper, a previous work in context-sensitive collocation identification using a sequence tagging BERT-based model improved with a graph-aware transformer architecture is used to investigate its feasibility to Japanese Language. The findings provide the initial insights into the automatic collocation typification in a non Indo-European language using deep learning models, and suggests that low resource languages can benefit from this approach.