Emotionally Aligned Responses through Translation
Abstract: Emotional response generation is an area of particular interest within conversational AI. However, many approaches lack control over the response. Potentially, due in part to the widely adopted approach of reflecting the users emotion in the response. As such this paper proposes an independent, but adaptable, emotion for the conversational agent that is separate from the user's, using Valence, Arousal, and Dominance scores which are updated based on user input. Additionally, by treating the alignment of the response as a matter of translation, a set of fine tuned sequence to sequence models are used to translate an initially generated response into one aligned with the agent emotion. This work provides a unique perspective on the topic of emotional response generation and showcases that potential means for improved consistency and controllability may yet be discovered beyond traditional methods.
Paper Type: Short
Research Area: Dialogue and Interactive Systems
Research Area Keywords: emotion detection and analysis, grounded dialog, conversational modeling, dialogue, text-to-text generation,
Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models, Theory
Languages Studied: English
Submission Number: 1117
Loading