Anatomy of an Idiom: Tracing Non-Compositionality in Language Models

TL;DR

This paper explores how transformer language models process idiomatic expressions by identifying specific neural circuits and attention patterns, revealing mechanisms for handling non-compositional language.

Contribution

It introduces novel techniques for circuit discovery in transformers and uncovers specific attention behaviors related to idiom processing.

Findings

Identification of 'Idiom Heads' attention patterns

Discovery of 'augmented reception' in idiom token attention

Insights into transformer mechanisms for non-compositional language

Abstract

We investigate the processing of idiomatic expressions in transformer-based language models using a novel set of techniques for circuit discovery and analysis. First discovering circuits via a modified path patching algorithm, we find that idiom processing exhibits distinct computational patterns. We identify and investigate ``Idiom Heads,'' attention heads that frequently activate across different idioms, as well as enhanced attention between idiom tokens due to earlier processing, which we term ``augmented reception.'' We analyze these phenomena and the general features of the discovered circuits as mechanisms by which transformers balance computational efficiency and robustness. Finally, these findings provide insights into how transformers handle non-compositional language and suggest pathways for understanding the processing of more complex grammatical constructions.