Aspect-level Text Sentiment Classification Based On Heterogeneous Graph Convolutional Networks

In recent years,with the rapid development of network technology,people are increasingly concerned about data.Massive data needs more efficient processing methods to filter.The task of text sentiment classification came into being.Aspect level text sentiment classification task is a major branch of text sentiment classification task,and it is also a hot issue in current research.There are two problems in the current research: firstly,most studies use the serialization research method,which can only extract the information near the aspect words,but can not obtain the long-distance information,while the existing isomorphic graph models only use the information between words and ignore other important information;secondly,the current model does not establish the relationship between aspects,which makes the aspects in the sentence unable to convey information to each other.To solve the above two problems,this paper proposes an aspect level text sentiment classification model based on heterogeneous graph convolution network.The main research contents are as follows: for the first problem,this paper sets up different nodes and edges to model the whole sentence,puts forward the structure of heterogeneous graph,sets up three different node types and five edge relationships.Each data in the dataset is mapped separately.The graph structure makes full use of the information of sentences.For the second problem,the edges between aspects are set in the graph,so that all aspect nodes in the sentence are connected,and the implicit relationship between aspects is mined.The heterogeneous graph convolution network model proposed in this paper consists of six parts.firstly,each data is segmented and corresponding to different nodes,which is represented by the initial word embedding vector,and the preliminary context representation vector structure is obtained through the bidirectional long-term and short-term memory neural network(Bi-LSTM);secondly,the graph representation matrix is input into the graph convolution network(GCN),and all nodes pass through the two-layer GCN network to obtain the vector representation with text features;thirdly,through aspect mask and attention layer,the final representation of relevant aspects is obtained;finally,the final classification result is obtained by the output layer.For sake of verifying the effectiveness of the algorithm model in this paper,four public data sets are used to verify it.Through experiments,it is proved that the four data sets are improved to varying degrees,which fully proves the effectiveness of the heterogeneous graph convolution network model proposed in this paper.