Deep language algorithms predict semantic comprehension from brain activity.

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Deep language models (DLMs) like GPT-2 excel at text processing tasks.
• The extent to which DLMs encode human-like comprehension remains debated.

Purpose of the Study:

• To investigate if GPT-2 representations align with human brain responses to language.
• To determine if these representations predict narrative comprehension levels.

Main Methods:

• Functional Magnetic Resonance Imaging (fMRI) data from 101 subjects listening to stories.
• Linear mapping model to predict brain activity from GPT-2 activations.
• Correlation analysis between model mapping and subject comprehension scores.

Main Results:

• GPT-2 representations significantly mapped onto brain activity.
• This mapping reliably predicted subjects' narrative comprehension.
• Key brain regions involved include angular, medial temporal, and supra-marginal gyri.
• Deep model layers capturing long-distance dependencies were most predictive.

Conclusions:

• DLMs like GPT-2 encode neural information relevant to human language comprehension.
• These models offer a valuable tool for understanding brain computations in language processing.
• Findings bridge artificial intelligence and neuroscience for language understanding research.