基于概念图卷积的方面级情感检测方法

Convolution over Syntactic Vocabulary Hierarchical Graph used in Aspect-Based Sentiment Analysis

摘要:

针对方面级情感分析方法准确率难以达到实用效果的问题,设计一种融合注意力机制并同时考虑句子句法结构和语料库共现信息的A-LSGCN模型,以便提高预测句子中特定属性情感极性的准确率；首先,联合多头注意力机制和词汇-句法图卷积,对属性的记忆向量和历史上下文内存向量进行叠加与更新,从而获得目标属性词及其上下文之间的关系；其次,为减少冗余对分类干扰,并充分学习通用语法知识,采用句法依存图神经网络提取句法结构信息,直接匹配属性及其情感表达,经网络分类计算最终得到特定属性对应的情感极性；最后在多个SemEval数据集上进行对比试验,其中Laptop14 数据集的MF1分数和准确率分别提升了1.1%、5.5%。

Abstract:

Abstract In order to solve the problem that the accuracy of aspect-based sentiment analysis method is difficult to achieve practical application, an A-LSGCN model that integrates attention mechanism and considers sentence syntactic structure and corpus co-occurrence information is designed, to improve the accuracy ,when predicting the sentiment polarity of each specific attribute in a given sentence. Firstly, combined the multi-head attention mechanism and the lexical-syntactic graph convolution for the superimposing of the attribute memory vector and the historical context memory vector, to obtain the relationship between the target attribute word and its contextual content. Secondly, in order to reduce the redundancy of the classification interference, and learn the general grammar knowledge enough, the syntactic dependency graph neural network is used to extract the syntactic structure information, directly match the attributes and their emotional expressions, and finally obtain the emotional polarity corresponding to the specific attributes through network classification calculation. In a comparative experiment on multiple SemEval datasets, the MF1 score and accuracy of the Laptop14 dataset improved by 1.1% and 5.5%, respectively.