MGCN: A Novel Multi-Graph Collaborative Network for Chinese NER

Named Entity Recognition (NER), one of the most important directions in Natural Language Processing (NLP), is an essential pre-processing step in many downstream NLP tasks. In recent years, most of the existing methods solve Chinese NER tasks by leveraging word lexicons, which has been empirically proven to be useful. Unfortunately, not all word lexicons can improve the performance of the NER. Some self-matched lexical words will either disturb the prediction of character tag, or bring the problem of entity boundaries confusion. Thus, the performance of the NER model will be lowered by such irrelevant lexical words. However, to the best of our knowledge, none of the existing methods can solve these challenges. To address these issues, we present a novel Multi-Graph Collaborative Network (MGCN) for Chinese NER. More specifically, we propose two innovative modules for our methods. Firstly, we build connections among characters to eliminate interferential influences of the noisiness in lexical knowledge. Secondly, by constructing relationship between contextual lexical words, we solve the problem of boundaries confusion. Finally, experimental results on the benchmark Chinese NER datasets show that our methods are not only effective, but also outperform the state-of-the-art (SOTA) results.