The Idiom Processing Advantage is Explained By Surprisal

Michaela Socolof, Timothy J. O'Donnell, Michael WagnerCogn. Sci 2025 · DOI

TL;DR

This paper investigates why we process idioms faster than literal phrases, proposing that "surprisal" – how unexpected a word is in context – is the key driver, not just compositionality. Through comprehension and production experiments, they found that lower surprisal in idioms largely explains the processing advantage, particularly for the noun component.

What Problem Does This Paper Address?

For decades, researchers have observed that people process idiomatic phrases (like "kick the bucket") faster than their literal counterparts. This "idiom processing advantage" has sparked debate: are idioms processed as single units (like words), or are they built compositionally from their parts? Existing theories often focus on "familiarity" or "compositionality" (how well the parts predict the whole). However, these measures have yielded mixed results, and there's no single, agreed-upon way to quantify compositionality. This paper aims to provide a more unified explanation by framing idiom processing through the lens of "surprisal theory," a well-established model in sentence processing that posits processing difficulty is driven by how unexpected words are in a given context.

Methods

The researchers conducted two main experiments:

  • Comprehension Experiment: Participants read sentences containing either idiomatic or syntactically matched literal verb-object phrases in a self-paced reading task. They measured reading times and analyzed the "surprisal" of individual words within the phrases. They also introduced a novel measure of "conventionality" to capture how predictable a word's meaning is in a specific context compared to its average meaning.
  • Production Experiment: Participants read sentences aloud, again containing either idioms or literal phrases. The primary measure here was spoken duration. They also analyzed surprisal and conventionality for the words in these phrases.

A key methodological choice was comparing idioms to syntactically matched literal phrases that don't have idiomatic interpretations. This helps avoid confounds from participants accessing unintended meanings.

Datasets

The study used custom-created sentence stimuli containing 17 different English verb-object idioms. These idioms were paired with syntactically matched literal verb-object phrases, either matching the verb or the noun of the idiom. Surprisal values were estimated using the GPT-3 language model, and conventionality scores were computed using BERT language model embeddings.

Key Insights and Ideas

  • Surprisal as the Unifying Factor: The paper strongly suggests that "surprisal" – the degree of unexpectedness of a word in its context – is a more fundamental predictor of processing speed than compositionality alone. Idioms, by their nature, often contain words that are less predictable in their specific idiomatic context compared to their literal counterparts, leading to lower surprisal.
  • Noun is Key: The idiom processing advantage, particularly in comprehension, seems to be driven primarily by the noun within the verb-object idiom. Nouns in idiomatic phrases showed significantly lower surprisal than nouns in literal phrases.
  • Compositionality's Nuance: While compositionality (measured as "conventionality") didn't directly explain the comprehension advantage, it showed a small role in production. This suggests that the way we understand and produce idioms might involve slightly different mechanisms.
  • Challenging the "Causal Bottleneck": The findings, especially the subtle role of conventionality in production, offer evidence against the strong version of surprisal theory where surprisal is a strict causal bottleneck for all processing factors.

Key Results

  • Comprehension: Idiomatic phrases were read faster than literal phrases. This advantage was largely explained by lower surprisal, especially in the noun component of the idiom. Conventionality did not show a significant effect on reading times.
  • Production: Idioms were spoken faster (shorter duration) than literal phrases, consistent with previous findings. This duration difference was primarily driven by the noun. Lower surprisal in idiomatic nouns significantly contributed to their shorter production times. Verb conventionality and its interaction with verb surprisal also had a small effect on verb duration, suggesting a more complex role in production.

Limitations and Future Work

  • Production Mechanisms: The paper acknowledges that the exact mechanisms driving the surprisal effect in production are not fully understood. They propose two possibilities: audience design (speakers adjusting duration for listeners) or planning facilitation (low-surprisal words aiding production planning). Further research is needed to distinguish these.
  • Generalizability: The current study focused on verb-object idioms. The authors suggest future work should explore other syntactic structures (e.g., adjective-noun idioms) to see if similar patterns hold and to better understand head-dependent asymmetries in processing.
  • Causal Bottleneck: While the results challenge the strict causal bottleneck of surprisal theory, they don't entirely dismiss it. The authors suggest that conventionality might not be immediately apparent during incremental comprehension but could play a role in production planning.

Summary updated

BibTeX
@article{socolof2025idiom,
  author = {Michaela Socolof and Timothy J. O'Donnell and Michael Wagner},
  title = {{The Idiom Processing Advantage is Explained By Surprisal}},
  journal = {Cogn. Sci.},
  year = {2025},
  doi = {10.1111/cogs.70085},
}
Figure