融合汉字字形信息的文本关系抽取

Text relation extraction integrating Chinese glyph

关系抽取任务旨在从文本中抽取实体对之间的关系,是当前自然语言处理领域的热门方向之一.由于中文具有复杂的句式和语法,导致现有的神经网络提取的特征以及语义表示能力较差,从而影响中文关系抽取的性能.汉字是象形文字,其字形结构在一定程度上隐含了字义,为此提出了包含字形级别实体表示的BERT_BIGRU_Glyph模型.模型中选用基于转换器的双向编码表征(BERT)为预训练模型、双向门控循环单元(BI-GRU)获取句子上下文表示.实体由字级表示和实体字形级表示共同构成,在实体字形表示中嵌入了BERT、BERT_CNN和BERT_BI-GRU三种提取字形特征的策略来丰富实体语义信息.实验结果表明:所提出的模型在实体字形相似的关系抽取中性能更优.

Relation extraction task is aimed to extract the relationship between entity pairs from text,which is one of the hot directions in the field of natural language processing.Due to the complex syntax and grammar of Chinese language,it leads to the poor feature extraction as well as semantic representation capability of existing neural networks,which affects the performance of Chinese relation extraction.Chinese characters are pictographs,and their glyph structure implies the word meaning to a certain extent.BERT_BI-GRU_Glyph model with glyph-level entity representation is proposed.In the model Bidirectional Encoder Representation from Transformers(BERT)is used as the pre-training,and Bi-directional Gated Recurrent Unit(BI-GRU)is used to obtain the sentence context representation.three glyph feature extraction strategies,BERT,BERT_CNN and BERT_BI-GRU,are embedded in the model to enrich entity semantic.Especially,entities consist of both word-level and glyph-level representations.The experimental results show that the performance of the proposed model is better in relation extraction with similar entity glyphs.