Named Entity Recognition Incorporating Chinese Segmentation Information

Abstract

Word-level information is crucial for Chinese named entity recognition. Presently, most works have achieved better performance by extracting word-level information into character-level representations through existing lexicons, but the maintenance of lexical lists is a major challenge. In this paper, we present the NIMSI model, proposing the incorporation of multiple segmentation information to enhance recognition, using a trilogy to align character-level attention with word-level attention to construct features of segmented information in Chinese text. Also, we use a simple but effective method to directly incorporate multi-segmentation information into character-level representations. Finally, as the experiments on the three benchmark datasets show, our model effectively incorporates segmentation information and alleviates the segmentation errors.