Invention Title:

GENERATIVE LANGUAGE MODEL SYSTEM WITH EMPATHY-TUNED EXPERT ADAPTER

Publication number:

US20260339933

Publication date:
Section:

Physics

Class:

G06F16/245

Inventors:

Assignee:

Applicant:

Smart overview of the Invention

The system enhances generative language models by incorporating empathy-tuned expert adapters. During the training phase, the system processes multi-turn conversations to identify empathy patterns across various tasks. It generates synthetic conversations reflecting these patterns, training an expert adapter to adjust the output of language models for specific empathy levels. This setup allows the system to generate responses that align with the desired empathy level for each task during the inference phase, improving user interaction satisfaction.

Generative language models typically produce responses with a uniform empathy level, which can lead to unsatisfactory user experiences, especially in contexts requiring varied empathetic responses. This limitation affects applications such as mental health support and customer service, where empathy is crucial. The system addresses this by integrating expert adapters and synthetic data to enhance the model's context-sensitive empathetic capabilities, improving the quality of AI interactions in diverse scenarios.

The implementation involves analyzing real-world conversational datasets to design empathy measurement methods. A dataset of 755 conversations across eight tasks revealed significant empathy differences before and after interactions. To counter data scarcity, a synthetic conversation generation pipeline was developed, steering responses towards defined empathy patterns. This approach ensures the generated responses closely match user expectations, enhancing the model's empathetic adaptability.

Expert adapters trained on synthetic conversations significantly reduce the gap between perceived and desired empathy levels. Results show a 72.66% reduction in this gap, with empathy scores improving by an average factor of 2.43 when expert adapters are used. These adapters maintain empathy patterns across conversation turns more effectively than system prompts, offering improved adaptability and user satisfaction in extended dialogues.

The system comprises processing circuitry, memory, and interfaces for user interaction. It clusters tasks based on empathy levels, identifying shared empathy clusters and patterns. Using a real-world dataset, the system demonstrated its ability to generate contextually appropriate empathetic responses across various tasks, enhancing AI assistants' effectiveness in delivering personalized, empathetic interactions.