Dependency-Aware Attention Model for Emotion Analysis for Online News

This paper studies the emotion responses evoked by the news articles. Most work focuses on extracting effective features from text for emotion classification. As a result, the valuable information contained in the emotion labels has been largely neglected. In addition, all words are potentially conveying affective meaning yet they are not equally significant. Traditional attention mechanism can be leveraged to extract important words according to the word-label co-occurrence pattern. However, words that are important to the less popular emotions are still difficult to identify. Because emotions have intrinsic correlations, by integrating such correlations into attention mechanism, emotion triggering words can be detected more accurately. In this paper, we come up with an emotion dependency-aware attention model, which makes the best use of label information and the emotion dependency prior knowledge. The experiments on two public news datasets have proved the effectiveness of the proposed model.