ABCD: A Simulation Method for Accelerating Conversational Agents With Applications in Aphasia Therapy

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Speech-language pathology

Background:

• Aphasia therapy development faces challenges like clinician shortages and patient recruitment.
• Existing methods for creating conversational AI for speech therapy are resource-intensive.

Purpose of the Study:

• Introduce Agent-Based Conversational Dialogue (ABCD), a novel method for simulating goal-driven dialogues between AI agents for aphasia therapy.
• Address barriers in therapy development using AI simulation, prompt engineering, and zero-shot techniques for cost-effective agent creation.

Main Methods:

• Developed AI clinician (Re-Agent) and AI patient (AI-Aphasic) using GPT-4o, integrated with speech-to-text and text-to-speech APIs.
• Simulated response elaboration training under varying semantic constraints (pictures, topics) and prompting techniques (Chain-of-Thought, zero-shot).
• Evaluated Re-Agent's performance using discourse metrics (coherence, grammaticality) against simulated aphasic errors at word and discourse levels.

Main Results:

• Re-Agent demonstrated accurate performance across all discourse metrics, semantic parameters, prompting techniques, and aphasic error levels.
• Zero-shot prompts yielded more direct, logically related responses robust to aphasic speech.
• Chain-of-Thought prompts occasionally resulted in minor local coherence loss due to complex reasoning.

Conclusions:

• ABCD provides a foundational computational approach to accelerate innovation and preclinical testing of AI conversational partners for speech-language therapy.
• This method circumvents the need for extensive, diverse errorful speech samples for AI fine-tuning.
• ABCD is designed to scale with advancements in AI and speech technologies, enhancing potential for clinical integration.