| Author: | Cheng, Yi |
| Title: | Towards emotional support conversational systems with goal awareness |
| Advisors: | Li, Wenjie Maggie (COMP) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Computing |
| Pages: | xviii, 177 pages : color illustrations |
| Language: | English |
| Abstract: | Emotional distress is a common haunting experience. Often, people cope with this distress by seeking emotional support through interpersonal interactions. However, emotional support from family and friends is not always available. To provide more people with timely emotional support, the development of Emotional Support Conversation (ESC) systems has gained significant attention. The rapid development in conversational AI, particularly those powered by sophisticated Large Language Models (LLMs), has made AI companionship increasingly plausible. Nonetheless, since LLMs are primarily optimized for passive instruction-following rather than goal-driven interaction, even state-of-the-art ESC systems built upon these LLMs can only respond to users' expression of distress in a reactive and echo-like manner in most cases. In contrast, effective emotional support demands goal awareness during conversation. A seasoned supporter must proactively explore the root causes of distress, strategically comfort the seeker's emotions, and guide them to determine how to improve the situation, all driven by a clear communication goal in mind. Without such goal awareness to proactively steer the conversation and gradually approach the dialogue goal, current ESC systems remain limited in providing effective emotional support. In this thesis, we identify the core research questions in building emotional support conversational systems with goal awareness, including: 1) Goal-driven Dialogue Planning: how to strategically plan the dialogue while considering the potential long-term effects of its interaction; 2) Dialogue Progression Analysis: as the dialogue progresses, how to monitor the dynamic dialogue progression (i.e., the extent of goal achievement) and further advance towards the dialogue goal; 3) Adaptation to Users: faced with users from diverse backgrounds, how to adapt to different users to fulfill the dialogue goal more effectively. This thesis provides a series of contributions aimed at addressing each of these fundamental questions. We introduce MULTIESC, an innovative ESC framework that performs goal-driven dialogue strategy over a long horizon. Unlike traditional approaches that conduct history-based dialogue planning, MULTIESC comprehensively considers each dialogue strategy's short-term and long-term effects, drawing inspiration from the A* algorithm that addresses the challenge of planning ahead by incorporating heuristic estimation of future cost. MULTIESC adopts novel lookahead heuristics to estimate the long-term user feedback after adopting a specific dialogue strategy by exploring a set of possible future dialogue trajectories. This approach advances goal-driven dialogue planning by considering how strategy choices influence the entire conversation in the long run, not just the next turn. Building on dialogue strategy planning, we further propose COOPER to address the challenge of monitoring dialogue progression when dealing with complex communication goals like emotional support, which are hard to measure in a quantifiable way. Grounded in the observation that complex dialogue objectives typically require the joint promotion of multiple dialogue goal aspects, COOPER coordinates a set of specialized agents, each tasked with managing a distinct aspect individually. By comprehensively analyzing the signals produced by the specialized agents, Cooper effectively monitors the dialogue progression and dynamically selects the goal aspect to prioritize during interaction. Finally, this thesis focuses on the crucial aspect of adaptation to users for long-term companionship, introducing SeaBench and AutoPal. Traditional systems depend on static user profiles or preset personas to tailor interaction, failing to adapt meaningfully over time as users' preferences and situations evolve. In this thesis, we take a step further and highlight the importance of autonomous, continuous adaptation to users over time, aiming for long-term companionship. We construct SeaBench, a comprehensive evaluation benchmark that assesses the foundational capabilities essential for such a self-evolving personalized conversational agent. Through extensive experiments, SeaBench exposes the limitations of current LLM-based agents in maintaining effective adaptation in long-term conversations. To address these limitations, we further develop AutoPal as a personalized agent for companionship that can autonomously adapt to the user's evolving needs through a hierarchical persona optimization framework. In summary, this thesis advances the development of emotional support AI systems with goal awareness, which are capable of proactive engagement, goal-oriented interaction, and personalized long-term companionship. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14067