A Novel Aspect-based Sentiment Classification Model With Multiple Grammatical Information

With the rapid development of the Internet and social network services,a large amount of text data is released on various platforms every day.These massive amounts of text data have stimulated the demand for text analysis by enterprises and governments.They have also provided sufficient impetus for the development of machine learning and promoted the vigorous development of the field of natural language processing.Sentiment analysis is a hot research direction in the field of natural language processing,while aspect-based sentiment analysis is a very challenging and widely used research problem in sentiment analysis.For example,an e-commerce platform can use aspect-based sentiment analysis to automatically extract consumers' attitudes towards different aspects of the product from product reviews;government can automatically sort out public attitudes towards social events by using aspect-based sentiment analysis on comments from online communities such as Weibo and Tieba;financial companies can automatically extract retail investors' prediction of stock trend through the discussion on social networks.Aspect-based sentiment analysis is quite difficult,and it can be divided into multiple sub-tasks,among which aspect-based sentiment classification(ABSC)has received great attention.The rapid development of pre-training language models has made breakthroughs in various research filed in natural language processing,and the application of graph convolutional network in natural language processing is also increasing.We propose BERT4 GCN,a model based on the BERT and graph convolutional neural network,for solving aspect-based sentiment classification.The research work of this paper includes:(1)We propose a model framework,BERT4 GCN,which uses both BERT and dependency tree as the source of grammatical information.BERT4 GCN introduces various types of grammatical information from the hidden states of the middle layers of the BERT and combine them with dependency-tree-based GCN,jointly considering grammatical sequential features and tree structure feature;Besides,we further prune and add the edge of the dependency graph based on self-attention weights in the transformer of BERT to deal with parsing errors and make dependency graph better suit ABSC task.In addition,we develop a method which incorporates relative positional embedding in node representations to make GCN position aware.Aspect based sentiment classification is a very challenging semantic understanding problem.In the situation of a small dataset,model is unable to learn general grammar knowledge,so external grammar knowledge can improve the effectiveness of the model.The BERT4 GCN model introduces a variety of grammatical information by using graph convolutional neural network based on dependency tree and BERT without using additional data.(2)We conduct experiments on three English datasets including Twitter,Sem Eval 2014Task4 Restaurant and Laptop.The result of experiments shows our BERT4 GCN model outperforms other prestigious aspect-based sentiment classification models.