Multi-channel Graph Convolution Models For Text Classification

Text classification is a basic and vital task in natural language processing.There are various application purposes,such as sentiment analysis,public opinion monitoring,news filtering,etc.Traditional convolutional neural networks and recurrent neural networks lack the ability to obtain non-continuous and long-distance semantic information,and cannot process non-linear structure data such as the semantic tree and syntax tree.Recently,the proposed graph convolutional text classification models can solve these problems effectively.However,the existing graph convolutional text classification models still have the following problems:(1)Since the existing spatial-based graph convolution models based on the single channel fail to effectively capture the implicit information such as the sentiment and syntax of the text,the ability of the model to obtain multi-directional text information is limited.(2)The existing models of sentiment classification task lacks the ability to learn the relationship between topics and sentiments.Since the graph convolution model is convenient to learn the information of different types of nodes on graphs with heterogeneous nodes,it is expected to learn the relationship between words,topics and sentiments,and further improve the performance of sentiment classification.In response to the above problems,the main research work of this paper is as follows:(1)A multi-channel spatial-based graph convolutional model for text classification is proposed.This model retains the original sequential-based graph.We use the cosine similarity to construct a semantic-based graph on the same corpus,and use a dependency parsing algorithm to construct a syntactic-based graph,so as to extract semantic and syntactic information from the text.In addition,we propose two feature fusion methods,simple fusion and attention fusion,to improve the encoding performance of the text representation.We test the performance of our methods on the three latest spatial-based graph convolution models of TextLevelGCN,TextING and MPAD.The experimental results show that the performance of the multi-channel spatial-based graph convolutional model is better than the original single-channel model,and the two feature fusion methods show their respective advantages on different types of datasets.(2)A dual-channel disentangled topic-document-word graph convolutional network is proposed.The DTPC-GCN model in controversy detection task has the ability to learn the relationship between the topic and the text,and can disentangle the topic-related features and the topic-unrelated features to learn the controversial category of the post.In this paper,we reconstruct the text graph of topic-document-word for the sentiment classification,and transfer the DTPC-GCN model to the sentiment classification task(we call it DTDW-GCN).Specifically,we utilize the LDA algorithm to generate topic distributions for the text,and use the TF-IDF algorithm and the PMI algorithm to construct a heterogeneous text graph containing nodes of topics,documents and words.Through multiple comparison experiments and ablation experiments,the results show that the DTDW-GCN model can effectively improve the performance of sentiment classification.On this basis,we further introduce semantic information and use the semantic similarity between word embeddings to construct a topic-document-word graph.We extend our dual-channel graph convolution method to DTDW-GCN,and realize the dual-channel model DCDTDW-GCN based on the multi-task learning framework.The experimental results show that the DCDTDW-GCN model has a significant performance improvement compared to the single-channel model.