Research On Sentiment Classification For Short Text

With the rapid development of information technology and the rapid popularization of mobile Internet,people are more willing to express their comments and share their lives on the Internet,which creates a large number of short text corpus containing rich emotional information on the Internet.Sentiment classification of short texts has long been a task not only in the academic sphere,but also in daily life,shopping websites can help consumers eliminate the information asymmetry between consumers and businessmen by analyzing consumption evaluation information,so that consumers can purchase satisfactory goods and enjoy high-quality services.Through the public opinion analysis system,the government uses short text analysis technology to guide the direction of public opinion,protect the interests of the people and maintain national security and stability.It has become an important means and way to serve the people in the new era.Therefore,the research of short text-oriented sentiment classification has important theoretical value and practical significance.In this paper,the topic of " Research on Sentiment Classification for Short Text " is studied,and the current research status and mainstream algorithms at home and abroad are introduced and analyzed in depth.Aiming at the difficulty of short text lacking abundant contextual semantic information,the following studies are carried out from two aspects: text representation and classification model.(1)Distributed model Paragraph Vector is an implicit semantic model.The meaning of every dimension of the vector trained by this model can not be explained,and the training of this model only uses the information of local windows,and can not make use of the information outside the window or even the whole corpus.To solve the above problems,this paper proposes a short text-oriented word-to-topic sentence vector model BTPV(Biterm Topic Paragraph Vector),which combines the global semi-dominant information from BTM(Biterm Topic Model)and the local implicit information from Paragraph Vector training process to train sentence vectors.The experimental results show that compared with the common distributed representation model,the short text based on this model achieves better clustering effect andprovides technical support for the research of short text emotional representation.(2)The Chinese sentiment classification method based on Word2 vec and LSTM needs to memorize the relationship between contextual words in the training process,which leads to its inefficiency and high cost.To solve this problem,this paper first proposes a distributed representation model BTSPV(Biterm Sentiment and Topic Paragraph Vector)based on the BTPV model in Chapter 3.The model integrates emotional information in the training process,and then proposes a sentiment classification method based on BTSPV and MLP.The experimental results show that compared with the methods based on Word2 vec and LSTM,the efficiency of this method is greatly improved.It achieves the relative balance of accuracy and efficiency required in practical use.Aiming at the task of short text sentiment classification,this paper systematically studies the difficulty of short text lacking rich contextual semantic information,and proposes a distributed representation model and sentiment classification algorithm for short text representation,which provides a new technical support for short text sentiment classification.