EmoChannelAttn: Exploring Emotional Construction Towards Multi-Class Emotion Classification

The current multi-class emotion classification studies mainly focus on enhancing word-level and sentence-level semantical and sentimental features by exploiting hand-crafted lexicon dictionaries. In comparison, very limited studies attempt to achieve emotion classification task from the emotion-level perspectives, which are to understand how the emotion of a sentence is constructed. Another limitation of existing works is that they assumed that emotion labels are relatively independent, neglecting the possible relations among different types of emotions. Therefore, in this work, we aim to explore various fine-grained emotions based on domain knowledge to understand the construction details of emotions and the interconnection among emotions. To address the first issue, we propose a novel method named EmoChannel to capture the intensity variation of a particular emotion in time series by incorporating domain knowledge and dimensional sentiment lexicons. The resulting information of 151 available fine-grained emotions is utilized to comprise the sentence-level emotion construction. As for the second issue, we introduce the EmoChannelAttn Network to identify the dependency relationship within all emotions via attention mechanism to enhance emotion classification performance. Our experiments demonstrate that the proposed method gains significant improvements compared with baseline models on several multi-class datasets.