Hierarchical Semantic Representations of Online News Comments for Emotion Tagging Using Multiple Information Sources

With the development of online news services, users now can actively respond to online news by expressing subjective emotions, which can help us understand the predilections and opinions of an individual user, and help news publishers to provide more relevant services. Neural network methods have achieved promising results, but still have challenges in the field of emotion tagging. Firstly, these methods regard the whole document as a stream or bag of words and can't encode the intrinsic relations between sentences. So these methods cannot properly express the semantic meaning of the document in which sentences may have logical relations. Secondly, these methods only use semantics of the document itself, while ignoring the accompanying information sources, which can significantly influence the interpretation of the sentiment contained in documents. Therefore, this paper presents a hierarchical semantic representation model of news comments using multiple information sources, called Hierarchical Semantic Neural Network (HSNN). In particular, we begin with a novel neural network model to learn document representation in a bottom-up way, capturing not only the semantics within sentence but also semantics or logical relations between sentences. On top of this, we tackle the task of predicting emotions for online news comments by exploiting multiple information sources including the content of comments, the content of news articles, and the user-generated emotion votes. A series of experiments and tests on real-world datasets have demonstrated the effectiveness of our proposed approach.