Research On Emotion Prediction Of News Articles From Reader’s Perspective

With the rapid development of Web2.0, network has been a public platform toedit and publish information. Everyone can share their opinions, moods throughnetwork which contains personal opinions, emotions and emotion feedback. And ithas important significance both in sociology and economics to indentify andmeasure the emotional change among people based on the change of time andtopics.In this paper, emotional feedback from the reader’s perspective is main objectfor study, to be specific, according to given news articles, analyze the words orsemantic information in the texts, and explore the reader’s emotions aroused by theaffective texts. At present, most studies on emotion analysis and detection focus onthe writer’s perspective, while the works from the reader’s perspective are relativelyless. What’s more, past studies primarily regard emotion prediction as a sinle-labelproblem, which think only one kind of emotion is generated, inconformity with thereality and the results of statistic. Besides these, most of existing works to predictemotions focus on Bog of words(BOW) model, however, according to psychologicalresearch, not only some specific words can arouse emotions, in many cases, reader’semotions are also associated with the events or topics of the texts. This studyresearches reader emotion prediction systematically under the network circumstance.Firstly, based on the study on the generation mechanism of reader emotions, makeuse of large-scale social votes to construct a multi-label reader emotion corpus. Andthe statistic of reader’s emotion votes shows one news article usually has multiplesignificant votes, so it is more reasonable to treat emotion pridiction as a multi-labelclassification task. In addition, the statistic also shows similar events often react tosimilar emotions, which means reader emotions are related to events/topics of text insome degree. Thus the study adopts BOW model and topic model respectively topredict reader emotions. And based on the original LDA(Latent Dirichlet Allocation)topic model, the advanced topic models are introduced to improve the performance,including the Weighted LDA and the Partitioned LDA. The final results of theexperiments on8,802news articles demonstrate that, multi-label classificationtechniques are more reasonable than those of the single-label on the study of readeremotional feedback. And, in the experiments of BOW model, the relative bettercombination can be achieved through adopting different classification algorithms toexperiment on different feature sets. Lastly, the topics models do well in predictingreader emotions. And through comparing the advanced topic models with theoriginal topic model, the proposed methods are shown effective and the average precision achieves0.89. In addition, the reader emotion corpus constructed in thestudy can used as important public resources to support related work.