Causal Drawbridges: Characterizing Gradient Blocking of Syntactic Islands in Transformer LMs

TL;DR

This paper investigates how Transformer language models process syntactic islands, revealing that they replicate human judgments and that certain mechanisms are selectively blocked, offering insights into linguistic representation.

Contribution

It introduces causal interventions to identify subspaces in Transformers responsible for syntactic processing, proposing a novel hypothesis about conjunction representation in linguistic structures.

Findings

Transformers replicate human judgments on syntactic island extraction.

Causal interventions reveal selective blocking of filler-gap mechanisms.

Conjunction 'and' is represented differently in extractable versus non-extractable constructions.

Abstract

We show how causal interventions in Transformer models provide insights into English syntax by focusing on a long-standing challenge for syntactic theory: syntactic islands. Extraction from coordinated verb phrases is often degraded, yet acceptability varies gradiently with lexical content (e.g., "I know what he hates art and loves" vs. "I know what he looked down and saw"). We show that modern Transformer language models replicate human judgments across this gradient. Using causal interventions that isolate functionally relevant subspaces in Transformer blocks, attention modules, and MLPs, we demonstrate that extraction from coordination islands engages the same filler-gap mechanisms as canonical wh-dependencies, but that these mechanisms are selectively blocked to varying degrees. By projecting a large corpus of unrelated text onto these causally identified subspaces, we derive a novel linguistic hypothesis: the conjunction "and" is represented differently in extractable versus non-extractable constructions, corresponding to expressions encoding relational dependencies versus purely conjunctive uses. These results illustrate how mechanistic interpretability can inform syntax, generating new hypotheses about linguistic representation and processing.