Improving Chinese Named Entity Recognition by Interactive Fusion of Contextual Representation and Glyph Representation

Named entity recognition (NER) is a fundamental task in natural language processing. In Chinese NER, additional resources such as lexicons, syntactic features and knowledge graphs are usually introduced to improve the performance of model. However, characters evolved from pictographs, their glyphs contain rich semantic information, which often ignored. Therefore, order make full use information contained character glyphs, we propose NER model that combines contextual representation glyph representation, named CGR-NER (Character–Glyph Representation for NER). First, uses large-scale pre-trained dynamically generate representations characters. Secondly, hybrid neural network combining three-dimensional convolutional (3DCNN) bi-directional long short-term memory (BiLSTM) designed extract glyph, potential word formation between adjacent global dependency sequence. Thirdly, an interactive fusion method with crossmodal attention gate mechanism proposed fuse different models dynamically. The experimental results show our achieves 82.97% 70.70% F1 scores on OntoNotes 4 Weibo datasets. Multiple ablation studies also verify advantages effectiveness