Automatic POS-tagging for Chinese Using Parallel Data

The aim of our study is to compare approaches to part-of-speech (POS) annotation and build a fine-tuned model for Chinese texts presented in the Russian-Chinese Parallel Corpus of the Russian National Corpus using parallel data. Automatic POS tagging is a fundamental task for natural language processing. Chinese poses a number of challenges for classical methods as it has little morphology and no spaces or other delimiters between words. State-of-the-art algorithms involve joint part-of-speech tagging and word segmentation. Another approach usually applicable to low-resourced languages suggests using the parallel data to improve labelling. Although the task of POS tagging for Chinese has been relatively well solved, texts translated from Russian contain a number of loanwords, which may cause a problem for traditional models. Therefore, in this work, we will apply the algorithms to such data and fine-tune the models using tagging of parallel Russian texts. For evaluation, we manually annotated test data, made a comparative analysis of the use of part-of-speech tags in different Chinese standards, and built a neural network based on parallel Russian-Chinese data.