Evidence Against Syntactic Encapsulation in Large Language Models.

Area of Science:

• Computational linguistics
• Artificial intelligence
• Cognitive science

Background:

• Large language models (LLMs) utilize attention mechanisms with specialized heads for linguistic tasks.
• A key question is whether syntactic computations in these heads are encapsulated or influenced by other information types.

Purpose of the Study:

• To investigate if semantic plausibility influences syntax-specialized attention heads in transformer-based LLMs.
• To determine if syntactic computations in LLMs are encapsulated or penetrable to semantic information.

Main Methods:

• Identified syntax-specialized attention heads in three LLMs (BERT, GPT-2, Llama 2).
• Tested the effect of semantic plausibility on attention weights within these specialized heads.

Main Results:

• Implausible semantic information reduced attention between words linked by syntax-specialized heads.
• This effect was observed across nearly all tested dependency types and models.

Conclusions:

• Syntactic information in LLM attention heads is penetrable to semantic plausibility.
• These findings suggest a integration of syntax and semantics, mirroring aspects of human cognition.