Think Twice: A Human-like Two-stage Conversational Agent for Emotional Response Generation

Abstract

Towards human-like dialogue systems, current emotional dialogue approaches jointly model emotion and semantics with a unified neural network. This strategy tends to generate safe responses due to the mutual restriction between emotion and semantics, and requires the rare large-scale emotion-annotated dialogue corpus. Inspired by the "think twice" behavior in human intelligent dialogue, we propose a two-stage conversational agent for the generation of emotional dialogue. Firstly, a dialogue model trained without the emotion-annotated dialogue corpus generates a prototype response that meets the contextual semantics. Secondly, the first-stage prototype is modified by a controllable emotion refiner with the empathy hypothesis. Experimental results on the DailyDialog and Empathet-icDialogues datasets demonstrate that the proposed conversational agent outperforms the compared models in the emotion generation and maintains the semantic performance in the automatic and human evaluations.

Conditioning on Dialog Acts improves Empathy Style Transfer

“…The availability of open-source datasets allowed the creation of many transformerbased empathetic response generation models (Li et al, 2020b;Majumder et al, 2020;Zheng et al, 2021;Sabour et al, 2022), all of which directly generate empathetic responses from the input prompt. Qian et al (2023) discovered that a two-stage system (response generation and style transfer) can yield better performance than one-stage models. However, a two-stage system requires two language models to be separately trained on different datasets.…”

Leveraging Large Language Models for Automated Dialogue Analysis

“…Human judges are commonly used when evaluating the degree of empathy exhibited in a dialogue response (Zhong et al, 2020;Sabour et al, 2022;Qian et al, 2023). There has also been some work on developing empathetic response and question taxonomies, although these are only applied in small-scale or synthetic settings (Welivita and Pu, 2020;Svikhnushina et al, 2022) (Zheng et al, 2021;Majumder et al, 2022) or for automatic evaluation (Kim et al, 2021;Lee et al, 2022).…”