MGSD: Multi-Graph Joint Framework Based on Semantic Dependency for Chinese NER

Abstract

In the realm of Natural Language Processing (NLP), Named Entity Recognition (NER) holds a pivotal position as a foundational component. Recently, the majority of existing approaches addressed the Chinese NER task by leveraging the lexicon enhancement method. However, these lexicon-dependent approaches can be susceptible to confusion caused by lexicon words, which can result in the recognition of false entities. Moreover, using only the lexicon in complex texts, dialects, and irregular sentences could lead to relatively poor results due to the lack of dependency information between Chinese words. To address these issues, in this paper, we propose a Multi-graph joint framework based on semantic dependency for Chinese NER (MGSD). We use semantic dependency relationships and lexical knowledge to construct four graphs that describe the connections between characters and words. After that, we leverage Graph Attention Network to extract features from these four graphs. With these features of Chinese phrases, our model can explicitly improve the issue of relying solely on lexicons. Experimental outcomes obtained from four Chinese NER datasets demonstrate the effectiveness of our model and outperform the state-of-the-art (SOTA) results.

The works described in this paper are supported by The National Natural Science Foundation of China under Grant Nos. 61772210 and U1911201; Key Research and Development Program of Guangdong of China under Grant No. 2023B0303010004; The Innovation Team Project for Universities in Guangdong Province in China under Grant No. 2023KCXTD011.