CSLAN: A Novel Lexicon Attention Network for Chinese NER

Rongsheng Lin, Shubin Cai, Zhong Ming

Abstract: Recently, many studies have employed lexicon-based models to integrate lexicon information into character repersentation for Chinese Named Entity Recognition(NER) tasks, and these models have demonstrated high effectiveness. However, previous approaches to incorporate lexicon information into character representations calculate the correlation solely between the current character and the matching words. They neglect the words matched by neighboring characters, leading to an underutilization of lexicon information in character sequences. To address this issue, we introduce a novel lexical attention network (CSLAN). This network improves character representation by incorporating a new lexical attention mechanism. This mechanism enables the current character to consider the lexicon information from neighboring characters when fusing it into the character representation, thereby regulating attention between the character and the associated vocabulary. We conduct extensive experiments, and results show that our proposed CSLAN achieves competitive results on four publicly available datasets.