Implicit Objective Network for Emotion Detection

Emotion detection has been extensively researched in recent years. However, existing work mainly focuses on recognizing explicit emotion expressions in a piece of text. Little work is proposed for detecting implicit emotions, which are ubiquitous in people's expression. In this paper, we propose an Implicit Objective Network to improve the performance of implicit emotion detection. We first capture the implicit sentiment objective as a latent variable by using a variational autoencoder. Then we leverage the latent objective into the classifier as prior information for better make prediction. Experimental results on two benchmark datasets show that the proposed model outperforms strong baselines, achieving the state-of-the-art performance.