Automatic Classification On Reader’s Emotion Towards News

With the rapid development of the internet, emotional text information released by the network users is undergoing a rather rapid expansion. The task of emotion classification aims to automatically classify the text into some emotional categories(e.g., happiness, anger, sadness, and fear). Recently, emotion classification has attracted many researchers’ attention in the research community of computational linguistics and becomes a basic hot topic. As a special kind of emotion, reader’s emotion specifically refers to emotion expressed by a reader after reading the text. This paper attempts to study reader’s emotion on news, and researches the following aspects of reader’s emotion classification:First, this paper proposes a respective coarse-grained emotion classification method based on co-training algorithm. The method is semi-supervised and its core idea is to jointly model reader’s emotion on news and writer’s emotion on comments to classify reader’s emotion. Specifically, we consider the news text and the comment text as two different views, and a co-training algorithm is then proposed to perform semi-supervised emotion classification. Experimental evaluation shows the effectiveness of our joint modeling approach.Second, this paper proposes a novel label propagation method to classify coarse-grained emotion. The method is semi-supervised and its core idea is to make full use of dependence between news and comments, and overcomes insufficiency of comments to perform semisupervised reader’s emotion classification. Specifically, we first construct two linking bipartite graphs, i.e., bipartite graph of the news-text samples and bipartite graph of the comment-text samples, to capture the dependence between the news text and the comment text. On this basis, we design a comment’s length-sensitive linear transition probability function to describe the emotional transfer strength between the news and comment text samples. Our two-view label propagation approach overcomes the limitations of co-training algorithm, and effectively overcomes the adverse effect on insufficiency of comments to improve the emotion classification results.Third, this paper proposes a feature-label factor graph model to classify fine-grained emotion. The method is supervised and its core idea is to joint feature space learning and label dependence modeling to perform multi-label reader’s emotion classification. Specifically, we generate several pseudo sample instances for each emotion label and build a pseudo sample network to represent the relationship among different emotion labels. On this basis, we propose a feature-label factor graph(FLFG) model to jointly learn the interaction between local textual features and relationship among emotion labels. Experimental evaluation shows that this method can effectively use label relationship to improve the performances.