Research On Sentiment Analysis Based On Topic Feature And Deep Learning

Text is the main medium for people to acquire knowledge and information.With the development of the Internet industry,the network has accumulated a large number of text data,such as news,comments,articles,etc..By mining and analyzing the subjective content of views,positions and emotions in these text data,people can quickly extract text information so as to obtain experience and make decisions.The features extracted from the text,the volume of which is increasing rapidly,often have the characteristics of high latitude.The traditional machine learning algorithm is difficult to deal with such complex data,which can only use language rules and dictionaries as auxiliary knowledge.In recent years,deep learning,which has been arose again,can gradually extract the deep features from the shallow features of the data,and almost do not need any auxiliary knowledge to deal with the high latitude data.The text analysis method of the deep neural network becomes the mainstream method.However,the current method of sentiment analysis lacks attention to the context of the text.This paper focuses on the emotional binary classification task of movie reviews,and proposes an emotion analysis method which combines text features and topic features.The main work of this article is:This paper studies and analyzes the basic flow of text sentiment analysis,including text preprocessing,text representation and building classification model.In text representation,in view of the problem of high dimension and semantic irrelevance,this paper uses the method of word2 vec word embedding to transform the text word into a vector containing a certain semantic information.In the process of building a classification model,this paper introduces the popular neural networks and the methods of building text analysis models.This paper improves the model of the bidirectional GRU neural network.The input layer of the network is constructed through vector stitching text features and document level theme features,and the attention mechanism is introduced after the input layer and the GRU layer.In this paper,the topic features of document level are considered in the task of sentiment analysis,and the attention of the network to the feature of useful text is strengthened through attention mechanism,while the attention to the useless features is weakened.The experiment shows that the performance of the original GRU neural network model is improved by theaddition of the topic feature and the addition of the double attention layer.The model of thispaper is better than the traditional neural network model in the movie review analysis task.From the perspective of multi-view learning,this paper proposes a co-training method basedon text features and text theme features.This method uses CNN and the improved GRUmodel as the two classifier for system training.The method of this paper is semi-supervisedlearning method,which reduces the dependence on labeled data.Experiments show that thismethod can effectively describe the role of text features and text topic features in sentimentanalysis,and achieve better classification results.