Leveraging Graph to Improve Lexicon Enhanced Chinese Sequence Labelling

Recently BERT has been employed for encoding a sequence of input characters in state-of-the-art Chinese sequence labelling models. However, Chinese sequence labelling often faces the lack of explicit word boundaries, which is well-noticed and more challenging problem. To alleviate this problem, we adopt the containing relation between characters and self-matched words from external lexicon to construct graph and incorporate lexicon-based graph information into the lower layers of BERT. We evaluate our model on ten Chinese datasets of three classic tasks containing Named Entity Recognition, Word Segmentation and Part-of-Speech Tagging. The experimental results demonstrate the effectiveness of our proposed method.