Aspect Level Sentiment Analysis Based On Attention Mechanism And Convolution Neural Network

With the gradual improvement of network infrastructure and the rapid development of mobile Internet,network platforms such as social software,shopping websites,and personal blogs have been welcomed by more and more Internet users.On these network platforms,users can freely publish emotionally colored commentary texts regardless of the limitations of time and space.These commentary texts describe their emotional tendencies towards certain hot events,public figures,commodities or services from various aspects,and have high social and commercial value.When users evaluate certain goods or services,they may express different emotional tendencies towards different aspects of the commodities or services.Aspect-level sentiment analysis,as a fine-grained task of sentiment analysis,aims to discriminate the user’s sentiment towards a specific aspect,and has extremely high application value.In recent years,deep learning technology has been widely used in the field of natural language processing.Using deep learning methods to perform aspect-level sentiment analysis in massive online text reviews has broad application prospects and has become a research hotspot in the field of sentiment analysis.This paper studies how to use attention mechanism and convolutional neural network to improve the effect of aspect-level sentiment analysis from three aspects: sentence semantic feature extraction,aspect word sentiment feature refinement and multi-feature joint modeling.Specifically,this research mainly includes the following three points:(1)Aiming at the defect that the Bi-LSTM-based aspect-level sentiment analysis method has insufficient ability to extract the global semantic features and local semantic features of sentences,this paper designs a multi-head self-attention convolutional neural network(Multihead Self-attention Convolutional Neural Network,MSCNN)aspect-level sentiment analysis method.Firstly,the multi-head self-attention mechanism is used to capture the long-dependent information of the sentence,to improve the global semantic feature extraction ability of the network,and to generate the aspect word representation in combination with the context where the aspect word is located.Furthermore,the multi-channel convolution module proposed in this paper is used to perceive the local information of sentences from multiple perspectives,and extract richer local semantic features.Finally,an attention mechanism is combined to generate aspect word sentiment features for classification.Experiments show that both modules in MSCNN can effectively enhance the semantic feature extraction ability of the network,thereby improving the performance of aspect-level sentiment analysis.(2)The traditional aspect-level sentiment analysis method based on attention mechanism ignores the syntactic-dependent features of sentences,which is easy to cause incorrect matching between aspect words and context,resulting in noise in the extracted sentiment features.To this end,this paper proposes an aspect-level sentiment analysis method based on the Sentiment Feature Refining Network(SFRN).Firstly,Bi-LSTM combined with attention mechanism is used to extract semantic features of aspect words and context,and GCN is used to extract syntactic dependent features related to aspect words.Furthermore,the emotional feature refining module based on the gating mechanism proposed in this paper is used to complete the semantic interaction and syntactic interaction between aspect words and context,and filter the noise caused by redundant information and incorrect matching in the emotional features of aspect words,so as to achieve better emotions.Analyze the effect.Experiments show that the accuracy and Marco-F1 value of SFRN on the three public datasets of Lap14,Rest14 and Twitter are better than the current mainstream methods.(3)Aiming at the problem that current aspect-level sentiment analysis methods ignore word co-occurrence information and syntactic label information,this paper proposes a Dual Strategies Graph Convolution Network(DSGCN)for aspect-level sentiment analysis.First,the semantic features of aspect words and contexts are extracted using Bi-LSTM.Furthermore,according to the syntactic dependencies between words in a sentence,a syntactic graph with syntactic labels and corpus-level word co-occurrence information are established to establish a word co-occurrence graph with word frequency labels.The graph convolution operation is performed on the graph and word co-occurrence graph,and the feature fusion is performed after extracting the syntactic information and word co-occurrence information of the sentence.Finally,after the emotional features are obtained by combining the attention mechanism,the Softmax classifier is used for emotional classification.Comparative experiments,ablation experiments and case analysis prove that DSGCN has positive significance for improving the effect of aspect-level sentiment analysis.