Reconstructing impaired language using generative AI for people with aphasia

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Healthcare
• Speech and Language Pathology

Background:

• Generative Artificial Intelligence (AI) and Large Language Models (LLMs) offer potential for assisting individuals with impaired language.
• Existing research on LLMs for aphasia often focuses on limited tasks, lacking real-world conversational reliability.
• Acquired communication disabilities, such as aphasia, significantly impact daily interaction.

Purpose of the Study:

• To develop and evaluate an AI-driven language-assistive solution for individuals with aphasia during natural conversations.
• To leverage LLMs for detecting and correcting speech errors, including neologisms, paraphasic errors, and word-finding difficulties.
• To enhance conversational fluency and coherence for adults with acquired communication disabilities.

Main Methods:

• Utilized Large Language Models (LLMs) with in-context few-shot prompting for error correction without model weight updates.
• Integrated LLMs into dialogue systems using the Langchain architecture to maintain conversational context.
• Trained and tested the system on a dataset of approximately 1980 utterances from 180 participants in the AphasiaBank corpus.

Main Results:

• The AI-reconstructed utterances achieved an 80% accuracy rate using the GPT-4o model.
• The system demonstrated the ability to correct neologisms, paraphasic errors, and word-finding gaps.
• Investigated the impact of various speech error types on the LLM's reconstruction accuracy.

Conclusions:

• LLM-based dialogue systems show significant promise for real-time language assistance in individuals with aphasia.
• The developed approach offers a more reliable solution for conversational support compared to task-specific LLM applications.
• Further research can refine the system by analyzing error types to improve correction capabilities for impaired speech.