A Context-Aware Gated Recurrent Units with Self-Attention for Emotion Recognition

Abstract

Text-oriented emotion recognition research has made significant research progress, however, very few works pay attention to learn the highquality long-distance contextual information for the utterance emotion recognition. In this paper, we proposed a context-aware gated recurrent units with self-attention for emotion recognition. The two bidirectional gated recurent units can obtain the sequence ralationship between words and utterances. Compared with the self-attention mechanism that cannot capture long-distance contextual information, three contexts are used in the self-attention mechanism, namely global, deep and deep-global context, which can capture the highquality long-distance contextual information. And the connection mechanism can enhance the embedding information of words and utterances. Experimental results on public Friends and EmotionPush datasets demonstrate that the three contexts in the self-attention mechanism outperforms several baselines on some emotion types recognition and indicate the effectiveness of the designed model, especially the deep-global context increased by 1% and 0.2% on UWA compared with all models.