Research On Reader’s Emotion Prediction Towards News Text

As the development of the Internet, the innovation on both type and content ofvarious kinds of social media is enhanced. One of the main point is to make theusers take their parts in it and share their own experience, which also make the largescale emotion analysis and tracking towards readers possible. The analysis andprediction of reader’s emotion towards media content could be used to improveuser’s information retrieval and reading experience, and also have large applicationon monitoring public opinion and other areas.Currently, the study in the emotion of text field mainly focus on the emotionanalysis, which means the emotion the text expressed. But the study on readers’emotion, which means the emotion caused by the text is still at the beginning. Thetopic of this paper is mainly on the method to predict the emotion reaction of readersafter reading news text, which means given a certain news report in a text form,reader’s emotion category prediction result could be given automatically through theanalysis of the news report text. The major work of this study includes, firstly, themethod towards news title is studied. Because the news title is short of words butwith various type of words and easy to cause a data sparseness problem, a methodwith the advantage of multilabel classifier and the semantic feature from HowNet isproposed. Based on the method above, the method based on the semantic sequenceis proposed. After the consideration of word order features the performance isimproved further; Secondly, as there isn’t any emotion dictionary resources foremotion prediction task, a new emotion dictionary based the separation of emotionexpression and emotion cognition is proposed. In this dictionary, emotion words isseparated between expression and cognition, thus could deal with both find-grainedemotion analysis and emotion prediction problems. Finally, a new emotionprediction method take the advantage of the dictionary mentioned above based onthe emotion type and intensity features is proposed, which is applied to both thenews main body text and title, and then combined with the method based onsemantic concept for news text to improve the prediction performance futher more.The experiment on the social tagging corpus’s result shows that the method based onthe semantic could decrease the feature dimension thus reduce the computation time,every performance evaluation could upgrade by more than6.8%. With the use ofannotation information, a relatively good result is achieved in a rather lowdimension feature set. It can be seen that the method proposed based on the semanticfeatures and human annotation knowledge could improve the performance of readeremotion prediction system significantly. More over, the new type emotion dictionary could be widely used in the emotion computation field.