Research On Fine-grained Sentiment Analysis Based On Memory Graph Convolutional Network

With the rapid development of the Internet and the improvement of user feedback system,fine-grained sentiment analysis plays an increasingly important role in the field of e-commerce and social media.As a key task in fine-grained sentiment analysis,aspect-level sentiment classification can determine the sentiment tendency of aspect words in the text.At present,researches on the aspect-level sentiment classification concentrate on the use of attention mechanism neural networks to focus on the sentiment words in comments by obtaining the importance of different words in the text sequence.However,these methods only obtain the sentiment information of the text and ignore the information related to the aspect words,which may lead to the problem of mismatching between the sentiment words and the aspect words.In order to solve the above problems,this paper proposes a memory graph convolutional network(Mem GCN)model which integrates the auxiliary information such as part of speech,aspect and position,and has the ability of syntactic perception.Firstly,the text representation layer integrates the part of speech and aspect information to assist the text representation,which can not only acquire the semantic and grammatical information embedded in the word,but also better express the character of the word in the task of sentiment classification.Secondly,the semantic learning layer uses the bidirectional gate recurrent unit to extract the abstract features of the context and to store the location information of the words in the location memory layer.Thirdly,the graph convolutional neural network takes the syntactic dependency tree of text as input,and learns the dependency relationship between words through syntactic information.Finally,the sentiment attention layer selects the sentiment words related to aspect words in the sentence,and assists the sentiment classification layer to judge the textual sentiment tendency.The objective function of Mem GCN model selects the cross entropy loss,and uses the accuracy and the F1 value of macro average as the main evaluation criteria of the model.We conduct extensive experiments on the Sem Eval-2014 and Twitter dataset.On Laptop and Restaurant datasets,the accuracy of the Mem GCN model is about 5～10% higher than that of the LSTM model,and the macro average F1 value on the Twitter dataset exceeds 70%.The experimental results show that the Mem GCN model can deal with the complex syntactic structure of aspect-level sentiment classification.At the same time,the Mem GCN model can solve the problem of mismatching between the sentiment words and the aspect words.There are 24 graphs,12 tables and 75 references in this paper.