Aspect Level Sentiment Analysis Based on Sentiment-Enhanced and Dual Graph Convolutional Network
Aspect Level Sentiment Analysis(ALSA)is designed to detect the sentiment polarity aspect of a given sentence.Most existing studies have reported the construction of Graph Convolutional Networks(GCN)on syntactic dependency trees to obtain syntactic information from aspect words and their contexts.However,this method has problems such as insufficient information extraction and a lack of sentiment information mining from sentences.To solve these problems,an ALSA model based on a sentiment-enhanced and dual-graph convolutional network is proposed.The model consists of a two-channel GCN that aims to mine the sentiment,syntactic,and semantic information in sentences.Position information and sentiment knowledge are used to construct a sentiment-enhanced dependency graph on syntactic dependency trees,and then a convolutional network of sentiment-enhanced graphs is constructed to enhance the sentiment-dependent relationship between aspect words and context,while simultaneously mining the rich syntactic information features in sentences.Subsequently,a GCN based on a multi-head attention mechanism is constructed to obtain the semantic information features in sentences.The output features of the dual GCN are masked,average-pooled,concatenated,and classified by the sentiment classification layer.The experimental results show that compared with the classical GCN model(ASGCN),the accuracy and F1 value of the Restaurant,Laptop and Twitter datasets improved by 3.43,5.69,3.13,3.92,3.57,and 4.02 percentage points,respectively.The proposed model has a better sentiment classification performance.
国家科技期刊平台
NSTL
万方数据
