Knowledge-enhanced Mixed-initiative Dialogue System for Emotional Support Conversations

Unlike empathetic dialogues, the system in emotional support conversations (ESC) is expected to not only convey empathy for comforting the help-seeker, but also proactively assist in exploring and addressing their problems during the conversation. In this work, we study the problem of mixed-initiative ESC where the user and system can both take the initiative in leading the conversation. Specifically, we conduct a novel analysis on mixed-initiative ESC systems with a tailor-designed schema that divides utterances into different types with speaker roles and initiative types.

we propose a knowledge-enhanced mixed-initiative framework (KEMI) for ESC, which retrieves actual case knowledge from a large-scale mental health knowledge graph for generating mixed-initiative responses.

Mixed initiative is commonly defined as an intrinsic feature of human-AI interactions where the user and the system can both take the initiative in leading the interaction directions (Allen et al., 1999; Kraus et al., 2020). For example, mixed-initiative conversational information-seeking (CIS) systems (Aliannejadi et al., 2019; Deng et al., 2023) can proactively initiate clarification interactions for resolving the ambiguity in the user query, instead of only reacting to the query. Accordingly, a mixedinitiative ESC system can proactively switch the initiative to provide an empathetic response or initiate a problem-solving discussion when appropriate.

Many efforts have been made on the emotion reasoning for generating empathetic responses (Shen et al., 2020; Zhang and Danescu-Niculescu-Mizil, 2020; Cheng et al., 2022; Peng et al., 2022). Another line of work focuses on identifying the dialogue acts of the utterances (Welivita and Pu, 2020; Malhotra et al., 2022; Svikhnushina et al., 2022) or predicting the next conversational strategies (Pérez- Rosas et al., 2017; Liu et al., 2021; Tu et al., 2022) in ESC systems. However, the feature of mixed initiative has not been investigated in existing ESC studies.

To facilitate the analysis on mixed-initiative ESC systems, we first propose an EAFR schema to annotate the utterances into different types with speaker roles and initiative types, named Expression (Userinitiative), Action (Support-initiative), Feedback (User Non-initiative), and Reflection (System Noninitiative). Besides, four emotional support metrics are designed to measure the characteristics of initiative and non-initiative interactions in ESC, including Proactivity, Information, Repetition, and Relaxation.

As shown in Figure 1, the system in ED solely targets at comforting the user by reflecting their feelings or echoing their situations, i.e., Non-Initiative. Differently, ESC systems are further expected to proactively explore the user’s problem by asking clarifying questions and help the user overcome the problem by providing useful information or supportive suggestions, i.e., Initiative. Furthermore, the analysis of the conversation progress and the emotional support metrics reveal three challenges in building a mixedinitiative ESC system: 1) When should the system take the initiative during the conversation? 2) What kind of information is required for the system to initiate a subdialogue? 3) How could the system facilitate the mixed-initiative interactions?

Furthermore, the analysis of the conversation progress and the emotional support metrics reveal three challenges in building a mixedinitiative ESC system: 1) When should the system take the initiative during the conversation? 2) What kind of information is required for the system to initiate a subdialogue? 3) How could the system facilitate the mixed-initiative interactions? According to these challenges, we define the problem of mixed-initiative ESC, which includes three sub-tasks: 1) Strategy Prediction to determine the mixed-initiative strategy in the next turn, 2) Knowledge Selection to collect the necessary knowledge for the next turn, and 3) Response Generation to produce emotional support responses with appropriate mixed-initiative strategy and knowledge. To tackle this problem, we propose a novel framework, named Knowledge Enhanced Mixed-Initiative model (KEMI), to build a mixedinitiative dialogue system for emotional support conversations with external domain-specific knowledge. In detail, KEMI first employs a knowledge acquisition module to acquire emotional support knowledge from a large-scale knowledge graph on mental health dialogues.