Research On Text Sentiment Analysis Based On Deep Learning And CTM Model

Since the beginning of the 21st century,the vigorous development of the Internet has promoted the birth of various e-commerce platforms.The emotional tendencies contained in the online review texts published on these platforms have a particularly important meaning,which can help users and businesses make effective decisions,and relying on manual methods to obtain information from these explosively growing texts is very time-consuming and laborious.Therefore,how to quickly and effectively mine valuable emotional information in massive texts has become a current research hotspot.Under this background,machine learning based on topic models and deep learning techniques based on neural networks have been successively applied to text sentiment analysis.Research field and made great progress.Most of the machine learning methods used in early research used the bag-of-words model to extract vocabulary features under the assumption that the topics are independent of each other,ignoring the correlation between topics,and using artificially designed features in feature selection.In recent years,with the rise of deep learning technology,the distributed word representation method has got rid of the problem of sparse word representation granularity in traditional machine learning,and has achieved good results.The problem of justice cannot be solved well.The improved pre-trained language model adopts the form of multiple sets of vector representations,which can solve the problem of polysemy,but in the model design,the a priori information about the meaning and relevance of the topic itself is not considered.In terms of sentence representation,there have been studies using a combined vector model representation method,but there is a defect that the sentence position information cannot be captured well.In extracting sentiment information of sentences,there are studies using multi-layer feed-forward neural network to extract sentiment of topic features,but this method has insufficient ability to extract hidden information.In order to solve the above problems,this paper combines deep learning technology with related topic models.First,this paper improves the deficiencies of traditional segmentation algorithms,and fully considers the correlation between topics:the CTM model is used to segment the topic features of the text,to obtain the correlation matrix between topics and between topics and words,and as the topic prior information input pre-trained language model;secondly,based on the topic prior information obtained by the text segmentation algorithm and the correlation vector of words and topics,the pre-trained ELMo model is used to dynamically express the text words,which can effectively solve the problem of polysemy,and The integration of the subject prior information plays a very good optimization effect on the subsequent information extraction.After that,on the basis of word representation,this paper will use the BiLSTM model to represent text sentences.It can consider the information before and after each word,and can better capture the sentence position information.Finally,it is integrated when extracting information from the sentence representation vector.The attention mechanism,which uses multi-head extraction to consider the global way,can find the hidden information in the text well,which is more sufficient than the pure multi-layer feedforward neural network extraction method,and can extract more comprehensive information of the text.The content of this article is mainly divided into five chapters.The first chapter introduces the background and current status of the study;the second chapter introduces the theory and development process involved in the study;the third chapter introduces the main idea of the topic segmentation algorithm used in this study in detail And compared with the traditional segmentation algorithm;Chapter 4 introduces the process of using deep learning to analyze the theme feature sentiment analysis and verify it with experiments;Chapter 5 summarizes this research and looks forward to the future research direction.