Dissociable Neural Mechanisms for Human Inference Processing Predicted by Static and Contextual Language Models.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Neuroscience

Background:

• Language models (LMs) demonstrate complex relationships with human language processing and neurophysiology.
• Recent studies show LM embeddings can create discourse representations for event inference.
• Understanding how different LM classes code event knowledge is crucial for mapping to human inference.

Purpose of the Study:

• Investigate event knowledge coding in distinct LM classes (static vs. contextual).
• Compare LM event inference capabilities with human experimental protocols (Metusalem et al., 2012; McKoon & Ratcliff, 1986).
• Explore the relationship between semantic processing and event inference across different LMs.

Main Methods:

• Utilized two established human experimental protocols for event inference.
• Compared inference performance of static LMs (word2vec, GloVe) and contextual LMs (BERT, RoBERTa).
• Assessed simpler semantic processing protocols alongside inference tasks.

Main Results:

• Static LMs showed a dissociation between semantics and inference for the two tasks.
• Contextual LMs exhibited a correlation between semantic and inference processing for both tasks.
• Static models performed Metusalem inference; contextual models succeeded in McKoon inference.

Conclusions:

• Inference tasks (Metusalem vs. McKoon) rely on dissociable processes.
• Contextual LMs' success in McKoon inference aligns with strategic human inference.
• Findings predict distinct neurophysiological markers for automatic vs. strategic inference processing.