面向中文的端到端多实体关系提取研究

End-to-End Multi-Entity Relationship Extraction for Chinese Text

关系提取可以获得文本中的关键信息。实体关系提取是在非结构化文本识别出实体并提取出实体对之间关系的方法。针对传统的关系提取借助外部NLP工具和局部分类等问题,端到端的实体关系提取模型可以减少管道模型之间的错误传播,获得更好的效果。设计一种基于BiLSTM的端到端的实体关系提取模型,通过决策对应实体和关系的可能性,解决中文文本中的多实体对的问题。在中文文本上,实现了3种实体识别10类关系提取。实验结果表明该方法无需人工构建复杂特征即可得到较好的多实体对关系抽取效果。

Relationship extraction can obtain key information in the text. Entity relationship extraction is a method to identify entities in unstructured text and extract the relationships between entity pairs. For the traditional relationship extraction with the help of external NLP tools and local classification, the end-to-end entity relationship extraction model can reduce the error propagation between pipeline models and obtain better results. An end-to-end entity relationship extraction model based on BiLSTM was designed to solve the problem of multiple entity pairs in Chinese texts by making decisions on the possibility of corresponding entities and relationships. In Chinese text, three kinds of entity recognition and 10 kinds of relation extraction are realized. The experimental results show that this method can achieve a better result of multi-entity relation extraction without manual construction of complex features.