Enhancing Emotional Support Capabilities of Large Language Models through Cascaded Neural Networks

Emotional support conversation aims to alleviate the emotional stress of the seeker and presents a challenging research task still in its nascent stage. It involves understanding the seeker’s emotional triggers and psychological intentions to provide suitable supportive responses. This can help mitigate negative feelings, especially when mental health resources are limited.While large language models like GPT have demonstrated powerful capabilities in areas such as writing, and long-text generation, aligning them for tasks in emotional support dialogue is problematic. These general-purpose models lack the logical reasoning and emotional comprehension required for complex human intentions, resulting in weaker emotional support capabilities.In this paper, we introduce a multi-model cascaded neural network system that combines dialogue state recognition, slot feature extraction, and generative language models. This framework enhances the emotional support capabilities of large language models by analyzing dialogue intentions and states, designing slots for different dialogue states, and extracting multi-dimensional information to guide the generative models. Our system utilizes non-deterministic finite state machines to model transitions between slot states, and we carefully design specific state transition functions to make the system more inclined to help users explore the deep-seated reasons for their emotions. The system can be configured to model existing psychological counseling and therapeutic methods, substantially improving the emotionally supportive text generated by large language models with only a minimal amount of training data.We evaluate the system’s capabilities in emotional support through multiple methods: automated assessments, human comparisons. Experimental results confirm that our system has a significant improvement in the ability of large language models to provide emotional support.