Detecting Idiomatic Expressions Through Cross Lingual Divergence

18 concepts9 papers1 proposal

When language is literal, it is compositional. A phrase like "the black cat" can be translated word-for-word into almost any language—such as le chat noir in French—while perfectly preserving both its structure and its core meaning.

However, when language becomes figurative, this compositionality breaks down. If you translate the English idiom "kick the bucket" word-for-word into French (couper le seau), the intended meaning of "to die" is completely lost. To preserve the meaning, a translator must map the phrase to an entirely different lexical item or expression.

This asymmetry means that cross-lingual translation can act as a diagnostic prism. By observing the statistical friction, structural shifts, and lexical mismatches that occur when phrases are pushed through translation models or parallel corpora, researchers can reliably isolate non-compositional and idiomatic language.

The Compositionality Gap

The most direct way to leverage translation for idiomaticity detection is to compare how a multi-word expression (MWE) translates as a whole unit versus how its individual components translate in isolation.

Imagine tracking the translations of the phrase "spelling bee." If the phrase is compositional, its target-language translation should closely match the combined translations of "spelling" and "bee." If it is idiomatic, the translation of the whole phrase will diverge drastically from the translations of its parts.

Compositional Phrase: "black cat"
[black] -> noir
[cat]   -> chat
[black cat] -> chat noir (High Overlap)

Idiomatic Phrase: "kick the bucket"
[kick]   -> donner un coup de pied
[bucket] -> seau
[kick the bucket] -> mourir (High Divergence)

Salehi et al. formalized this intuition in Detecting Non-compositional MWE Components using Wiktionary. By evaluating the lexical divergence between an expression's overall translation and its component translations across multiple languages, the authors created a highly effective unsupervised predictor for non-compositionality.

This concept of translational friction has deep roots in statistical machine translation. Melamed demonstrated in Automatic Discovery of Non-Compositional Compounds in Parallel Data that treating a sequence of words as a single token within a translation model significantly increases the model's predictive power if—and only if—the phrase functions as a non-compositional semantic unit.

Expanding this to a massive scale, Pichotta et al. developed a "polyglot ranking" method in Identifying Phrasal Verbs Using Many Bilingual Corpora. By measuring how much a verb's translation probability distribution shifts when it is paired with a particle (e.g., "take" vs. "take off") across 50 different languages, they successfully mapped out idiomatic phrasal verbs with human-level accuracy.

Round-Trip Translation Friction

Another way to expose an idiom is to force it through a round-trip translation (back-translation). If you translate a literal sentence into a pivot language and then translate it back to the source language, the original meaning and structure usually emerge intact. But if a sentence contains an idiom, the round trip often introduces severe lexical degradation.

If a machine translation system translates an idiom literally into the pivot language, the figurative meaning evaporates. When the pivot text is translated back, it will surface as a bizarre literal phrase rather than the original idiom.

Tiedemann et al. leveraged this exact behavior in Helsinki-NLP at SemEval-2022 Task 2: A Feature-Based Approach to Multilingual Idiomaticity Detection. They showed that discrepancies between an original phrase and its back-translated counterpart serve as a powerful zero-shot signal for identifying figurative usage in text.

To make deep learning models explicitly aware of this phenomenon, Yayavaram et al. integrated round-trip translation directly into neural network optimization in BERT-based Idiom Identification using Language Translation and Word Cohesion. The authors designed a custom loss function that penalizes models based on semantic shifts during back-translation, quantified by metrics like METEOR. This forces the underlying language model to pay closer attention to the boundaries of non-literal expressions.

This strategy of using back-translation to probe linguistic boundaries dates back to early rule-based systems. Nitta introduced the Cross Translation Test (CTT) in Idiosyncratic Gap: A Tough Problem to Structure-bound Machine Translation, comparing structure-preserving literal translations with meaning-preserving free translations to systematically expose cultural and idiomatic gaps between languages.

Repurposing Machine Translation Failures

Machine translation models are traditionally optimized to assume compositionality. Consequently, their inherent struggle to handle figurative language can paradoxically be repurposed as a diagnostic feature.

Hauer et al. utilized this insight in UAlberta at SemEval 2022 Task 2: Leveraging Glosses and Translations for Multilingual Idiomaticity Detection. They demonstrated that the systematic failure of literal word-for-word translation to preserve meaning provides a highly resilient signal for classifying idiomaticity, especially in low-resource settings where large annotated datasets are unavailable.

Similarly, Boitet et al. turned this limitation into a classification tool in Towards an Automatic Classification of Illustrative Examples in a Large Japanese-French Dictionary Obtained by OCR. By translating dictionary examples into a common third language (English), the authors showed that ordinary, literal usage results in high lexical overlap across translations, whereas proverbs and idioms generate massive lexical divergence. The machine translation model's confusion becomes the exact feature used to classify the text.

Complementing Translation with Lexical Cohesion

While cross-lingual divergence exposes how an expression behaves externally, idiomaticity also leaves an internal footprint within the source language: it disrupts local semantic harmony.

In a literal phrase like "kick the ball," the words share a strong contextual relationship; "kick" and "ball" frequently co-occur in sports contexts. In the idiom "kick the bucket," the semantic link between the physical act of kicking and a household bucket completely vanishes. The idiom introduces a sudden drop in local semantic cohesion.

Ehren explored this internal signal in Literal or idiomatic? Identifying the reading of single occurrences of German multiword expressions using word embeddings, demonstrating that measuring the weakening of lexical cohesion via word embeddings can effectively determine whether a specific instance of an MWE is being used literally or figuratively.

Ultimately, the most robust modern architectures, such as the one proposed by Yayavaram et al., combine both perspectives. By pairing internal lexical cohesion metrics with external cross-lingual translation loss, models can evaluate a phrase from both sides: identifying the internal semantic disruption within the sentence, and verifying it via the inevitable friction that occurs when it crosses language barriers.

Key Foundational Work

Automatic Discovery of Non-Compositional Compounds in Parallel Data (1997) — contributed 1 concept elements to this direction.
Idiosyncratic Gap: A Tough Problem to Structure-bound Machine Translation (1986) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_lingual_translation_divergence_92c8f8d7,
  title = {cross-lingual translation divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-translation-divergence/rdr_92c8f8d7bc7143f49b6a54af0ad21a48},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_lingual_translation_divergence_92c8f8d7_v1,
  title = {cross-lingual translation divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-translation-divergence/rdr_92c8f8d7bc7143f49b6a54af0ad21a48/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake