Generating Completions for Broca's Aphasic Sentences Using Large Language Models

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Speech and Language Therapy

Background:

• Broca's aphasia presents with non-fluent, effortful speech and good comprehension.
• Current aphasia treatments are often time-consuming and lack real-world conversational relevance.

Purpose of the Study:

• To explore the application of sequence-to-sequence Large Language Models (LLMs) for sentence completion in Broca's aphasia.
• To assess LLMs' efficacy in reconstructing agrammatic speech patterns.

Main Methods:

• Generated synthetic Broca's aphasic data using a rule-based system.
• Fine-tuned four pre-trained LLMs on sentence completion using synthetic data.
• Evaluated model performance on both synthetic and authentic Broca's aphasic data.

Main Results:

• Demonstrated LLMs' capability in reconstructing agrammatic sentences.
• Observed improved model performance with longer input utterances.
• Showcased potential for LLMs without requiring authentic aphasic samples for initial training.

Conclusions:

• LLMs show significant potential for improving communication aids for individuals with Broca's aphasia.
• The approach may be applicable to other clinical populations with speech impairments.
• LLM-based interventions offer a promising, scalable alternative to traditional therapies.