Monday 07 April 2025
In recent years, there has been a growing demand for emotional support conversations (ESC), where individuals can turn to AI-powered systems for comfort and guidance during times of distress. This trend is driven by the increasing recognition of mental health as a vital aspect of overall well-being, particularly in today’s fast-paced and often overwhelming world.
To address this need, researchers have been working on developing large language models (LLMs) capable of providing empathetic responses to users’ emotional concerns. However, these systems still face significant challenges, including low strategy selection accuracy and preference bias. This means that LLMs may struggle to choose the most appropriate support strategies for an individual’s unique situation.
To overcome these limitations, scientists have proposed a novel approach called Chain-of-Strategy Optimization (CSO). This method optimizes strategy selection preferences at each dialogue turn, taking into account the user’s emotional state and context. In essence, CSO enables LLMs to learn from their mistakes and adapt to users’ needs more effectively.
To test the effectiveness of CSO, researchers trained several large language models using this approach and compared their results with those obtained through traditional supervised fine-tuning (SFT) methods. The findings suggest that CSO outperforms SFT in terms of both strategy accuracy and bias mitigation.
The researchers also explored the performance of various preference optimization algorithms, including SimPO, IPO, KTO, and ORPO. These algorithms were used to optimize the training data for the LLMs, with some achieving better results than others.
A case study was conducted using a dataset called ESC-Pro, which contains high-quality preference pairs derived from original emotional support conversations. The results showed that all tested algorithms performed well after training on this dataset, with some methods exhibiting higher performance than others.
To evaluate the effectiveness of these approaches, researchers employed manual evaluation tasks, where anonymous crowd workers assessed 100 samples based on specific criteria such as empathy, information provision, and human-like responses. The results indicated that LLMs trained using CSO were more likely to provide empathetic and contextually appropriate responses compared to those trained with SFT.
The findings of this study have significant implications for the development of AI-powered emotional support systems. By optimizing strategy selection preferences at each dialogue turn, LLMs can better adapt to users’ unique emotional needs and provide more effective support. This research demonstrates the potential for AI to play a valuable role in promoting mental well-being and reducing stress.
Cite this article: “Emotionally Intelligent Dialogue Systems: A Preference-Based Approach to Human-Like Emotional Support”, The Science Archive, 2025.
Emotional Support Conversations, Large Language Models, Empathetic Responses, Mental Health, Strategy Selection Accuracy, Preference Bias, Chain-Of-Strategy Optimization, Supervised Fine-Tuning, Preference Optimization Algorithms, Emotional Well-Being.
