Research Of Target-dependent Sentiment Analysis Based On Hierarchical Neural Network

With the rapid development of the information technology and the advent of a new era of the Internet,the daily lives of modern people are closely related to the Internet.More and more people have changed because of e-commerce platforms and social network platforms.At the same time,people are on the Internet leaving lots of comments.How to effectively analyze the emotional tendencies has a very important meaning,sentiment analysis has come into being.Sentiment analysis is based on mining a large number of Internet texts to analyze the emotions of the opinion holders.The work of sentiment analysis mainly focuses on document-level sentiment analysis,sentence-level sentiment analysis and aspect-level sentiment analysis.As document-level and sentence-level sentiment analysis are coarse-grained research methods,they ignore the fact that sentiment expression is often attached to a specific target,leading to bias in sentiment prediction.Therefore,the article studies target-dependent sentiment analysis.The paper mainly made three improvements.Firstly,research on text semantic classification method based on hierarchical neural network,through the document-level sentiment analysis to achieve the purpose of text semantic classification.Secondly,research on target-dependent sentiment analysis based on the hierarchical neural network is conducted to study sentiment tendencies by learning target information.Target-dependent sentiment analysis research is more effective and better classified than the research on text semantic classification.Thirdly,the research on sentiment analysis based on hierarchical neural network for word-aspect fusion and positional encoding is carried out,and the research of sentiment analysis is carried out by in-depth study of the objective aspect information and introduction of positional encoding.The works of the paper mainly include the following contents:Firstly,research on text semantic classification method based on hierarchical neural network.The hierarchical attention network model(HBLSTM-ATT)for document-level sentiment classification is proposed.The method encodes the semantics of sentences and the relationship with document representation,performs sentiment analysis and training.A large number of experiments on the four datasets show the effectiveness of the method.Secondly,research on target-dependent sentiment analysis based on hierarchical neuralnetworks.The target-dependent sentiment analysis based on hierarchical neural network model(HTAS-BiGRU)is proposed.The key idea of ??the model is to learn the target information of sentences so as to capture the sentiment of sentences and improve the performance of target-dependent sentiment classification.In addition,the improved hierarchical attention mechanism module includes the target-words attention module and sentiment classification attention module.A large number of experimental results at the sentence level and document level show that the model in the chapter achieves the best target-dependent sentiment classification performance.Thirdly,research on hierarchical context attention network with word-aspect fusion and positional encoding based on hierarchical neural network.The hierarchical context attention network with positional encoding(HCAN-PE)model is proposed.The proposed model in the chapter consists of hierarchical content attention mechanism,which includes word-aspect attention mechanism layer and sentence attention mechanism layer.The hierarchical context attention mechanism is mainly presented how to build the document representation progressively from word vectors hierarchically.In order to make full use of the order and arrangement of words in sentences,hierarchical positional encoding is added.The Word-Aspect Fusion Attention is introduced to tackle the difficulty of model training about the word-aspect na?ve concatenation.Compared with several baseline models,qualitative experiments were performed on the two datasets,and the experimental results achieve the state-of-the-art state.At the same time,a detailed visualization study of attention mechanism is done.