Extending deliberate degradation of an artificial neural language model to induce dementia-like deficits

Anonymous

Abstract: Recent advances in speech and language technologies aim to leverage clinical information embedded in a person's language abilities to automatically assess cognitive health and function. In this work, we investigate possible perturbations of large language models that could lead to behaviors compatible with those observed in clinical conditions. In particular, we perturb GPT-2 to observe the impact on a generation task used to assess Alzheimer's dementia (AD). Our work achieves statistically significant degradation of the model, and additional classification experiments demonstrate that lexico-syntax is the most impacted linguistic apparatus during deliberate degradation of GPT-2. These findings could inform diagnostic pathways and medical interventions of AD.

Paper Type: long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Contribution Types: NLP engineering experiment

Languages Studied: English