Research On Twitter Sentiment Classification Based On Sentiment Word Embedding And Convolutional Neural Networks

With the advent of the Big Data and the rapid development of social network applications,people are increasingly willing to share their feelings or publish their views on micro blogging,such as Facebook,Twitter and other social media,and they also give comment on the product or service they bought in E-commerce website.Analyzing this kind of subjective text efficiently and accurately can help the company and government make the right decision.Therefore,it is necessary for us to study the sentiment analysis on this text.In this paper,we study the short text sentiment analysis task and select twitter text as the main research object,specifically,we output the emotional polarity including positive?negative and neutral for a given tweet text.We first build a multi-granularity emotion-enhanced word embedding model(also called MEWE in short)that can learn semantic and emotional information from both word level and sentence level.The MEWE model makes full use of distant supervised corpus with emotional polarity and sentiment lexicon,finally the model produces word embeddings that contain both semantic relations and rich emotional information.In addition,to extend sentiment word embeddings from words to text,we introduce the deep learning technology to model the text and propose a sentiment classification system based on Convolutional Neutral Network and multi-feature fusion.We input the multi-granularity emotion-enhanced word embeddings produced by MEWE to our build CNNs model and obtain the deeper and implicit text features.As for the feature fusion,since there are many well-designed manual features that help a lot for sentiment classification in existing work,and the short text is limited by its length and lack of context environment,simply using the word embedding as features to classify may not be effective.So we try to concatenate the depth sentiment word embedding generated by the CNNs model with the manual features and then use this merged feature to predict the emotional polarity of text.The final sentiment classifier does not use the softmax regression in CNNs directly,but use SVM model instead.Especially for three class classification problem,we design the One-Versus-One SVM to conduct one-to-one model training and classification on positive,neutral and negative texts and finally get the emotional polarity.In order to verify the validity of our method,we designed various sentiment analysis experiments:firstly,we conducted word level experiments to evaluate MEWE model which we analyzed the semantic similarity of the word and measured the accuracy of the word sentiment classification quantitatively;Then,we experimented on sentence-level sentiment classification to verify the effectiveness of the CNNs and multi-feature fusion based approach and carried out the three class classification and two class classification on SemEval's twitter corpus.Furthermore,we also compared the benchmark system and others approach with our method.The experimental results show our method can solve the sentiment analysis task effectively.