融合多层注意力机制与双向LSTM的语义关系抽取

Multi-level Attention-based Bidirectional Long Short-Term Memory Networks for Relation Extract

关系抽取是构建如知识图谱等上层自然语言处理应用的基础。针对目前大多数关系抽取模型中忽略部分文本局部特征的问题,设计一种结合实体位置特征与多层注意力机制的双向LSTM网络结构。首先根据位置特征扩充字向量特征,并将文本信息向量化,然后将文本向量化信息输入双向LSTM模型,通过多层注意力机制,提高LSTM模型输入与输出之间的相关性,最后通过分类器输出关系获取结果。使用人工标注的百科类语料进行语义关系获取实验,结果表明,改进方法优于传统基于模式匹配的关系获取方法。

Relational extraction is the basis for constructing upper natural language processing applications such as knowledge graph.Because most of state-of-the-art systems ignore the importance of the local feature,in this paper,we design a bidirectional LSTM network structure which combines position eigenvector and multi-level attention mechanism.Firstly,the model embedded the text information by extending word vector feature which was based on positional features.Secondly,the information was introduced into the bidirectional LSTM model,and multi-level attention was used to improve the probability between the input and output of LSTM model.Finally,it obtained the result by classifier.The proposed method in this paper achieves an better result than the tradition method.