基于深度神经网络和注意力机制的实体关系抽取方法研究

Entity Relation Extraction Based on Deep Neural Network and Attention Mechanism

摘要: 关系抽取是自动构建知识图谱的关键技术之一，其根本目标是抽取实体间的语义关联关系。针对非结构化文本实体关系抽取方法中存在的上下文环境信息难以准确表征、句子间的实体关系特征未被充分利用等问题，本文提出了一种新的基于神经网络和注意力机制相结合的关系抽取模型PCNN-ATT-BiLSTM。该模型采用“RNN + CNN”网络框架，其中的RNN利用双向长短期记忆神经网络(Bi-LSTM)来捕获文本语句的上下文信息和浅层语义特征，利用分段卷积神经网络(PCNN)捕获文本语句的局部短语特征，并结合注意力机制捕获文本语句的关键信息进行关系预测。该模型在公开数据集SemEval-2010 Task8上取得了82.92%的F1值，实验结果表明，该方法在非结构化文本的实体关系抽取方面表现出了较好的性能，为实体关系的自动获取提供了新的方法支持。

Abstract: The relation extraction is one of the key technologies to automatically construct knowledge graph, and its fundamental goal is to extract semantic relation between entities. The current methods for extracting entity relation from unstructured text are difficult to accurately describe the context information, and the entity relation between sentences is not fully utilized. This paper proposes a new relation extraction model PCNN-ATT-BiLSTM based on the combination of neural network structure and attention mechanism. The model adopts the “RNN + CNN” framework, in which RNN uses the Bi-directional Long Short-Term Memory network (Bi-LSTM) to capture the context information and shallow semantic features of text data statements, uses the Piecewise Convolutional Neural Network (PCNN) to capture the local phrase features of text statements, and combines the attention mechanism to capture the key information of text statements, so as to better capture the information in sentences for relationship prediction. The model achieves 82.92% F1 value on the public dataset SemEval-2010 Task8. The experiment shows that the method shows good performance and provides a new method support for the automatic acquisition of entity relation.