基于双向GRU神经网络和双层注意力机制的中文文本中人物关系抽取研究

CHARACTER RELATION EXTRACTION IN CHINESE TEXT BASED ON BIDIRECTIONAL GRU NEURAL NETWORK AND DUAL-ATTENTION MECHANISM

实体关系抽取是知识抽取的重要组成部分。与传统模式识别的方法相比,深度学习的方法在关系抽取任务中表现得更为突出。目前关于中文的关系抽取技术的研究主要是基于核函数和远程监督的方法,而且数据集中的噪音数据对实验结果带来的负面影响不可忽视。提出一种基于双向GRU神经网络和双层注意力机制的中文关系抽取模型。结合中文语言的结构特点,采用字向量的形式进行输入,针对遗忘性问题,采用双向的GRU神经网络对输入向量进行融合。从一个句子中提取出字级别的特征信息,并通过句子级别的注意力机制来提取句子特征。利用远程监督的方法在新闻网站上抽取约8 000条数据进行验证。实验结果表明,双层注意力机制的神经网络模型可以充分利用句子的所有特征信息,准确率和召回率相较于未加入注意力机制的神经网络模型都有显著提升。

Entity relationship extraction is an important part of knowledge extraction.Compared with the traditional methods,the deep learning is more prominent in relation extraction tasks.The current research on Chinese relation extraction techniques is mainly based on kernel functions and remote supervision methods.Moreover,the negative impact of noise data in data sets on experimental results cannot be ignored.This paper proposed a Chinese relation extraction model based on bidirectional GRU neural network and dual-attention mechanism.In combination with the structural characteristics of the Chinese language,the input was in the form of a word vector.Because of the forgetting problem,we used a bidirectional GRU neural network to combine input vectors.We extracted word-level feature information from the sentence and the sentence feature from sentence-level attention mechanism.We used remote supervision methods to extract about 8 000 data on the news website for verification.The experimental results show that the neural network model of the dual-attention mechanism can make full use of all the feature information of the sentence,and the accuracy and recall rate have been further improved compared with the neural network model without attention mechanism.