GAMJL: A Chinese Named Entity Recognition Method Based on Gated Attention Mechanism and Joint Learning
Yi Wang, Junwu Zhu, Erfeng Xu, Wei Lin · 2024
Chinese text, lacking clear word boundary delimiters, poses challenges for Named Entity Recognition (NER). Existing methods extract character sequence features using external dictionaries or deep learning models, but often fail to filter shared features, which may lead the model to misjudge entity boundaries and incorporate irrelevant semantic information. This paper introduces a Chinese NER method based on a gated attention mechanism and joint learning. The model first encoded Chinese text using a pre-trained language model XLM-RoBerta, then obtained entity features and entity span features through a bidirectional Long Short-Term Memory network (Bi-LSTM) and linear transformation layers. A gated attention mechanism was then constructed to filter a nd f use t he i nput f eatures. T he fused features significantly i mproved t he m odel’s a bility t o perceive entity types and spans. Experimental results on the Resume and MSRA public datasets showed that the proposed method outperformed other baseline models.