Topic Models For Social Emotion Detection

With the fast development of online social platforms,more and more users are no longer confined to obtain information from the Internet and turn to express their opinions and emotions.Sentiment analysis refers to the use of natural language processing,text analysis,computational linguistics to systematically identify,extract,quantify,and study affective states and subjective information.Sentiment analysis is widely applied to reviews,survey responses,online and social media,and clinical medicine.Social emotion detection,a task under sentiment analysis,aims to detect readers' emotions towards a news article.Emotions of a large quantity of uses can shape into public opinion and attitude,which has dramatically practical significance.In the existing methods,the discriminant model ignores the topic information implied in the article and can only obtain one classification result,which cannot be used to analyze the cause of the emotion.Generating models(for example,topic models)usually make bag-ofwords assumption,ignoring the order of words,and consider that the topic and emotion of each word in the document are independent,and this oversimplification has both advantages and disadvantages.Therefore,this article focuses on the use of topic models for social sentiment detection,and is committed to abandoning the bag-of-words assumption and considering the relationship of words in documents.The main work of this paper is as follows:1.The study of introducing the sentence structure of documents and the relationship between adjacent sentences into topic models for social emotion detection was conducted.We propose the Topic-Emotion Transition Model that combines a hidden Markov model with a topic model.The model takes into account sentence structures and the transition of topics and emotions between adjacent sentences.The model can be used for sentiment analysis at both the document level and sentence level simultaneously and considers the correlation between emotions.The sentiment analysis results of the TET model outerform the state-of-the-art methods at both document and sentence level on two evaluation metrics.2.The study of introducing concept into topic model and simultaneously modeling emotion,event category,and topic for social emotion detection was conducted.We first propose a concept extraction method based on dependency parsing,and propose the Emotion and Event Category Discovery model(EECD).The basic unit of processing in the EECD model is concept.Concepts contain the co-occurrence of words and the words can be used to associate concepts with semantic relevance.And then based on the EECD model,a generalized pólya urn model was adopted to further propose the EECD-GPU model which considers the semantic relevance of concepts via word embeddings.EECD and EECD-GPU can be used to analyse topic,sentiment and event category of documents and analyse the event categories that trigger readers' emotions.The results of social emotion detection conducted on news headlines outperform the state-of-the-art methods.The paper consists of five chapters.The first chapter introduces the research background and significance,current research status,research motivation and goals.The second chapter introduces the topic model for sentiment analysis and other related technologies.The third chapter introduces the social emotion detection method and experiment based on TET model.The fourth chapter introduces the social sentiment detection methods and experiments based on EECD and EECD-GPU models.The fifth chapter is the summary of the work of the paper and the prospect of future work.