反馈机制的实体及关系联合抽取方法

Entity and Relationship Joint Extraction Method of Feedback Mechanism

实体及关系抽取是信息抽取中的两个核心任务,是构建知识图谱的重要基石。对于实体识别和关系抽取,当前通常采取人工提取特征和规则,分独立两步实现的方法,这种方法易造成数据重复预处理和错误传播。实体识别和关系抽取两个模块存在相互关联性,实体识别是进行关系抽取的基础,实体关系抽取结果又可反馈校验实体信息。因此,文中提出无须添加人工特征和引入互反馈机制的混合神经网络模型(Mufeedback-Join Model)来完成实体及其关系的联合抽取,实现实体关系对实体识别的反馈校验机制。该模型共享Bi-LSTM特征提取层来提取文本上下文特征,依据共享层特征引入Attention机制捕获关键局部特征来完成解码,再用条件随机场CRF完成实体序列的标注任务,融合共享层特征和实体特征,并将其输入到CNN模型来实现实体关系的抽取,最后计算关系抽取结果的损失值,再联合实体识别损失值反馈修正特征提取层和实体识别模型参数。将此算法应用在实体数据集上进行实验,在同等硬件环境下,该方法可以缩短的模型训练时间,提升实体及关系抽取的准确率、召回率和F1值,联合抽取的F1值整体提升了3.91%,实体识别子模块的F1值平均提升了1.34%,关系抽取的F1值提升了5.79%。实验数据说明,联合抽取模型可以实现两个子模块的合并,从而缩短数据处理时间和减少错误数据的传递;相互反馈的机制可以提升整体识别效果。

Entity and relationship extraction are two core tasks in information extraction,and are the important cornerstone of knowledge mapping.At present,entity recognition and relationship extraction usually adopt the method of extracting features and rules manually and realizing them independently in two steps.This method is easy to cause duplicate data preprocessing and error propagation.The two modules are interrelated.Entity recognition is the basis of relationship extraction.The results of entity relationship extraction can also feedback and verify entity information.Therefore,a hybrid neural network model(Mufeedback-Join Model)without adding manual features and with mutual feedback mechanism was proposed to extract entities and their relationships jointly and realize the feedback checking mechanism of entity relationship to entity recognition.The model shares Bi-LSTM feature extraction layer to extract text context features,and introduces attention mechanism to capture key parts based on shared layer features.After decoding the feature,CRF is used to complete the entity sequence labeling task.The shared layer feature and entity feature are input into CNN model to realize entity relationship extraction.Finally,the relationship extraction result loss value is calculated,and the feature extraction layer of loss value feedback correction and the parameters of entity recognition model are combined.In the same hardware environment,this method can shorten the training time of model,improve the accuracy,recall and F1 value of entity and relationship extraction,The F1 value of the joint extraction is improved by 2.91%,the entity identification sub-module F1 is increased by 1.34%on average,and the relationship extraction F1 value is increased by 5.79%.The experimental data show that the joint extraction model can merge two sub-modules to reduce data processing time and error data transmission,and the mechanism of mutual feedback can improve the overall recognition effect.