融合多注意力神经网络的方面级情感分析  

Aspect-based sentiment analysis based on multi-attention neural network

作　　者：梁燕[1,2] 刘超 梁仲雄 李文涛 LIANG Yan;LIU Chao;LIANG Zhong-xiong;LI Wen-tao(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Key Laboratory of Signal and Information Processing of Chongqing,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学信号与信息处理重庆市重点实验室,重庆400065

出　　处：《计算机工程与设计》2023年第3期894-900,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61702066);重庆市教委科学技术重点研究基金项目(KJZD-M201900601)。

摘　　要：针对现有方面级情感分析模型中特征向量信息不足和语义丢失问题,提出一种多注意力融合的神经网络模型。利用词向量注意力机制捕捉句子上下文和方面词之间的语义关系;利用位置注意力机制影响方面词周围的情感特征;利用自注意力机制捕捉序列内部特征用于加强序列表示。为验证模型的有效性,在SemEval 2014 Task 4和ACL 14 Twitter基准数据集上进行实验,实验结果表明,所提模型取得的性能优于比较方法。Aiming at the problems of insufficient feature vector information and semantic loss in the existing aspect level emotion analysis models,a neural network model of multi attention fusion was proposed.Word vector attention mechanism was used to capture the semantic relationship between sentence context and aspect words.The affective features around aspect words were influenced using positional attention mechanism.The self-attention mechanism was used to capture the internal features of the sequence to enhance the sequence representation.To verify the effectiveness of the model,experiments were carried out on SemEval 2014 Task 4 and ACL 14 Twitter benchmark data sets.The experimental results show that the performance of the proposed model is better than that of the compared methods.

关 键 词：深度学习 循环神经网络 自然语言处理 方面级情感分析 注意力机制 词向量 位置编码 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]