Research On Social Media Text Personalized Sentiment Analysis Based On Matrix Factorization

Social media is a kind of new online interactive platform,which has greatly changed the way of people's communication and thinking,and results in exploration of user generated information.The large amounts of social media texts generated by users have become one of the most representative big data sources.Mining and analyzing the user generated information has great influences on social development.As a kind of information processing technique in analyzing,preprocessing,summarizing and inferring subjective text,social media text sentiment analysis has recently drawn great attention from research and industry communities,and has wide applications on the Internet.Traditional text sentiment analysis research mainly focuses on analyzing sentiment from texts,but ignores individual differences on sentiment expression,which weakens analysis quality.To solve the above questions,this dissertation is about the research on social media text personalized sentiment analysis.As matrix factorization techniques have been widely used in social media processing,several variants are proposed in this dissertation to tackle possible challenges in social media text personalized sentiment analysis.Specifically,the main research work and contributions are listed as below:(1)To address the problem of few reliable and large-scale fine-grained sentiment lexicons(or emotion lexicons),this work proposes a joint non-negative matrix factorization method which constructs a fine-grained sentiment lexicon from news corpus,and then applies it to text sentiment analysis.Based on crowd-sourced multi-label news corpus,this method factorizes text-word and text-label matrices simultaneously into several topic-related matrices;and then combines them into a fine-grained sentiment lexicon via compositional semantics method.This work compares the sentiment lexicons built by different methods and verifies the effectiveness of the proposed lexicon construction method.Meanwhile,the constructed lexicon is also verified to have better quality and usability than those existing state-of-the-art fine-grained sentiment lexicons.(2)For the problem of individual differences on sentiment expression,a personalized microblog sentiment classification model is developed based on the latent factor model,which decomposes texts into words to solve the sparsity of personal data,considers following relations between users to encode similar interests for capturing personalized preferences more effectively,and introduces syntactic units derived from the dependency relations between words as text features for modeling fine-grained text semantics.All of these will enable the model to predict personalized sentiment from microblogs more accurately.The experimental results based on real-world Chinese weibo and English tweet datasets show that the proposed model can capture the influences of individual differences more effectively than traditional sentiment analysis models.(3)For the cold-start problem in personalized text sentiment analysis,this work considers both user and product information and proposes a text-driven latent factor model for product review sentiment analysis.In order to address the cold-start user and cold-start product problems,a pairwise rating comparison optimization strategy PRC is proposed to learn better user and product feature parameters based on rating comparison information in reviews,and achieves improvements on overall performance.Experimental results on three public review datasets show that the proposed method is much better than state-of-the-art neural network method in both sentiment analysis quality and learning efficiency.(4)For the problem of global users and products influences on text semantics,this work considers the advantages of both matrix factorization and deep neural network,and then proposes a deep matrix factorization model.Firstly,this method uses a multi-layer perceptron component to capture high-level representation of user-product interactions;then utilizes convolutional neural network component to model local semantics of various granularities;and finally combines both the two representations into the sentiment score based on a generalized matrix factorization component.Experimental results based on three public review datasets show that the proposed method is much better than traditional sentiment analysis methods and collaborative filtering recommendation methods.(5)As there are few available real-time personalized sentiment analysis systems,this work is based on above achievements and implements an available and real-time personalized microblog sentiment analysis prototype system.The system provides user-friendly charts to analyze the sentiment orientation of microblogs published by login users and their followees,including sentiment tendency prediction and sentiment proportion analysis of microblog users,which provides sentiment knowledge supports for user actions such as browsing,reply,retweet and like.In summary,this work focuses on the research of social media text personalized sentiment analysis based on matrix factorization,which includes fine-grained sentiment lexicon construction,personalized sentiment analysis on microblogs,personalized sentiment analysis on product reviews,cold-start problem and real-time sentiment analysis prototype system.The research achievements lay theory foundations and provide technique supports for constructing and implementing social media text personalized sentiment analysis.