HADE: Hierarchical Affective Dialog Encoder for Personality Recognition in Conversation

Anonymous

Abstract: Personality recognition in conversation aims to determine the personality traits of speakers through the dialogue content, which is of great importance in designing personalized conversational AI. Existing methods that use only linguistic patterns in utterances limit their performance. To fill in the gap, we investigate the effectiveness of incorporating affective information and modeling the interactions among speakers in conversations for personality recognition. However, available corpus with personality and explicit affective annotations is rare. Besides, modeling the dialog flow with multiple speakers is difficult. Faced with the issues, we proposed Hierarchical Affective Dialog Encoder (HADE) for effective personality recognition in conversation. HADE utilizes manual annotated Valance-Arousal-Dominance (VAD) vectors of single words and implicitly extracts affective information from utterances. Then, it introduces a hierarchical architecture with the dialog state embeddings to identify the speakers and encode the whole dialog flow. Finally, the affective information is integrated by an auxiliary VAD regression task to enhance personality recognition. Extensive experiments on a well-known dataset, \textbf{FriendsPersona}, demonstrate the effectiveness of our method compared with state-of-the-art models. Besides, we conduct an ablation study to discuss different approaches for integrating affective information and dialog flow modeling; the design of both parts in HADE is also verified to be effective for personality recognition in conversation.