The Study On Text Sentiment Prediction With Sentiment Emphasis And Sentiment Fusion

Nowadays,with the rapid development of the Internet,people can publish texts about specific objects on various platforms.Different from the numerical data,these texts can reflect subjective opinions and express the sentiment or emotional attitude of the publishers,making the text itself to be the most common sentimental catharsis tool.Therefore,conducting sentiment analysis on texts and mining the underlying sentiment or attitude,has great theoretical significance and application value in decision-making,service adjustment,public opinion monitoring and other aspects.However,unlike other common statistical features,text sentiment can't be obtained directly,and the current works on processing text sentiment mainly conducted via classification,or the form of text sentiment processing is relatively simple.Therefore,this thesis focuses on the "text" as well as its sentiment,adding the factor of sentiment into different text processing scenarios,so as to introduce a new"dimension" to be considered by the conventional text processing methods.This thesis mainly focuses on the following issues:First,we devote effort to the text sentiment polarity detection based on the word co-occurrence.Such the problem considers the word as a minimum unit.By ascertaining word's sentiment polarity,we can effectively classify the sentiment of a given text as positive or negative.Secondly,we explore the self-played text integration with sentiment emphasis.This problem aims to integrate a large quantity of comments on the given object into a short integrated text,and to enhance the sentiment-intensity of the integrated text.Thirdly,we study the text sentiment prediction in a multi-participant communication context with sentiment fusion.This problem analyzes the changes on sentiment of the historical texts posted in a multi-participant communication context,so as to predict the sentiment of text to be published in the future.In addition,the main innovations of this thesis are as follows:1)Propose a two-aspect lexicon expansion method,which can revise the sentiment polarity of the words(already included in the lexicon)and infer the sentiment polarity of words(excluded in the lexicon)according to the current text-context,so as to overcome the drawback of word's enumeration and the ambiguity of word's sentiment polarity.According to the experimental results,compared with the existing methods,detecting text sentiment polarity based on the above two-aspect lexicon expansion can achieve much better precision;2)Conduct text integration in a self-play way which means only the source documents are required in order to fine-tune the model and generate the integrated text.In this way,the data cost for obtaining the "gold standard" on the integrated text can be avoided.In addition,during the self-playing,we additionally introduce the knowledge on sentiment-emphasis so that the integrated text can have higher sentiment intensity objectively,no matter the sentiment is positive or negative.According to the experimental results,combining self-play and sentiment-emphasis mechanism can achieve the overall most optimal performance on content quality and sentiment intensity in terms of the generated integrated text;3)By adopting the principle of time series prediction,it can be guaranteed that the sentiment of text,to be published in the future,could be predicted,even if the text-content is still unknown;Moreover,in this thesis,we select a target-participant from the multi-participant communication context,and before sentiment prediction,the sentimental stimulation on the target-participant,triggered by the other participants,will be captured first.And then,the sentiment of texts posted by the target and other-participants will be fused as a complete consideration on the final prediction.According to the experimental results,incorporating sentiment fusion into the time-series based sentiment prediction can overall bring about the most optimal precision on the text sentiment prediction.Based on the above works,finally,we combine and adjust above critical methods as a complete sentiment prediction method with sentiment emphasis and sentiment fusion.Then,apply such an "adjusted" method to the sentiment prediction task on the movie monthly review.According to the experimental results,compared with the existing methods,combining and adjusting the critical methods in this thesis,with an application on the movie monthly review,it can achieve a comparatively obvious effect on sentiment-emphasis as well as an overall better precision on the text sentiment prediction.