Towards Emotional Support Dialog Systems

TL;DR

The paper defines Emotional Support Conversation (ESC) and a three-stage ESC Framework grounded in the Helping Skills Theory, then builds ESConv, a richly annotated dataset with supporter training and rigorous quality control. Through transformer-based backbones and strategy-conditioned variants, it demonstrates that incorporating explicit support strategies improves both automatic metrics and human judgments of emotional support quality. Key contributions include the ESC task formalization, the ESConv dataset with detailed annotations, and empirical evidence that strategy-aware generation enhances empathetic, exploratory, and action-oriented support in dialogue. The work lays a data-and-framework foundation for developing practical, peer-level emotional support dialog systems while addressing data quality and ethical considerations.

Abstract

Emotional support is a crucial ability for many conversation scenarios, including social interactions, mental health support, and customer service chats. Following reasonable procedures and using various support skills can help to effectively provide support. However, due to the lack of a well-designed task and corpora of effective emotional support conversations, research on building emotional support into dialog systems remains untouched. In this paper, we define the Emotional Support Conversation (ESC) task and propose an ESC Framework, which is grounded on the Helping Skills Theory. We construct an Emotion Support Conversation dataset (ESConv) with rich annotation (especially support strategy) in a help-seeker and supporter mode. To ensure a corpus of high-quality conversations that provide examples of effective emotional support, we take extensive effort to design training tutorials for supporters and several mechanisms for quality control during data collection. Finally, we evaluate state-of-the-art dialog models with respect to the ability to provide emotional support. Our results show the importance of support strategies in providing effective emotional support and the utility of ESConv in training more emotional support systems.

Figures (8)

• Figure 1: An example chat showing effective emotional support (adapted from ESConv) being provided to the help-seeker(left) by the supporter(right). The support strategies (skills) used by the supporter are marked in the parentheses before the utterances. The red bold texts in the dashed boxes highlight the three stages of our proposed ESC Framework (Figure \ref{['fig:framework']}).
• Figure 2: Emotional support conversations (our work) can include elements of emotional chatting zhou2018emotional and empathetic respondingrashkin-etal-2019-towards.
• Figure 3: Overview of our proposed ESC Framework. It contains three stages and suggested support strategies. The procedure of emotional support generally follows the order: ①Exploration$\to$ ②Comforting$\to$ ③Action (as indicated by the black arrows), but it can also be adapted to the individual conversation as needed (indicated by the dashed gray arrows). The column of "Lexical Features" displays top 5 unigrams or bigrams associated with messages that use each strategy in our dataset. Each feature is ranked by the rounded $z$-scored log odds ratios monroe2008fightin in the parentheses.
• Figure 4: The distribution of strategies at different conversation progress.
• Figure 5: The Joint model's generation distribution. The meanings of all the graphics and abbreviations are consistent with Figure \ref{['fig:distribution']}.