Sentiment Analysis Method Based On Joint Learning Of User-word Representation

Internet has deep influence on human society.Nowadays people surf the Internet for not only uploading and downloading information but also communicating with each other for abundant information and closer sentiment connection.This makes it possible to observe massive user sentiment expresssion.How to analyse users’ sentiment expression on Internet and mine the latent information is not only an important requirement from industry development but also a key problem in scientific research.Sentiment analysis is an important topic in natural language processing.The most exisitng works in this area focus on the text classification using textual features without the consideration of the effect of human beings in sentiment analysis.Such a problem influcences the performance of sentiment analysis.This study carried out two parts of work towards this problem.Firstly,an heterogeneous network,which contains both users and words,is constructed.The representation of users and words are jointly learnt through network nodes embedding.In this way,users and words are completely separated while the users and words of different polarities are separated.Secondly,the learned user and word representation from heterogeneous network nodes are used as the input of a convolutional neural network based classifier to construct a sentiment analysis model.The experimental results on IMDB,Yelp2013,Yelp2014 datasets show 2.6%,1.8% and 1.4% improvement are obtained,respectively.This study proposed a joint learning method of user and word representaion by means of incorporting user information in the process of heterogeneous network nodes embedding.This method increases the descriptive ability of sentiment analysis model,and thus improves the performance text sentiment analysis.The idea of joint learning of user and word representations may be applied to other representation learning problems.Moreover,this study applied linguistics theory to guide the design of deep learning algorithms.This brings some innovation in the research of this area.