Evaluating scientific theories as predictive models in language neuroscience.

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Data-driven encoding models excel at predicting brain responses to language.
• However, these models lack interpretability, failing to explain which stimulus features drive neural activity.
• A gap exists between qualitative scientific theories and quantitative data-driven models.

Purpose of the Study:

• To introduce Question Answering (QA) encoding models for interpretable prediction of brain responses to language.
• To convert qualitative theories of language selectivity into quantitative, predictive models.
• To bridge the gap between scientific theories and data-driven neuroscience.

Main Methods:

• Utilized large language models (LLMs) to answer yes-no questions based on qualitative theories, annotating language stimuli.
• Developed compact QA encoding models (using 35 questions) for predicting brain responses.
• Validated models using functional Magnetic Resonance Imaging (fMRI) and Electrocorticography (ECoG) data.

Main Results:

• A compact QA encoding model significantly outperformed existing baseline models in predicting brain responses.
• The model demonstrated high accuracy in both fMRI and ECoG datasets.
• Interpretable model weights provided quantitative maps of language selectivity across the cortex.

Conclusions:

• QA encoding models effectively translate qualitative theories into accurate and interpretable models of brain responses.
• LLMs can bridge the gap between theoretical neuroscience and data-driven predictive modeling.
• The developed method offers a powerful tool for understanding the neural basis of language processing.