---
title: "Idioms & multiword expressions — papers"
description: "Idioms & multiword expressions: 141 papers of the Elamur Research library, the most cited first."
canonical: "https://research.elamur.ai/free/en/topics/multiword-expressions/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/multiword-expressions/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Idioms & multiword expressions

# 🧷 Idioms & multiword expressions

141 papers · page 1 of 6 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Phrases that mean more than their words: idioms, collocations, light verbs and phrasal verbs.

💡 Explained: [Lexicography](https://md.research.elamur.ai/free/en/concepts/lexicography/)

A survey · about 20 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

A multiword expression is a group of words that a speaker stores and uses, at least in part, as a single unit. The class is wide. It takes in opaque idioms such as *kick the bucket*, phrasal verbs such as *look up* or *turn off*, light-verb constructions such as *take a shower* or *make a decision*, compound nouns such as *traffic light*, fixed function phrases such as *according to*, and conventional collocations such as *strong tea*, where nothing is strange about the meaning but the word choice is fixed by habit. What unites them is that some property of the whole, whether its meaning, its syntax, or simply its frequency, cannot be predicted from the words taken one at a time.

For computational linguistics these expressions are a standing embarrassment. A system that treats text as a sequence of independent words will translate *spill the beans* as a remark about legumes, parse *look the number up* as a verb with a stray preposition, and assign a WordNet sense to *red* in *red tape*. A system that instead lists every expression as a fixed string will miss *the beans were finally spilled*, and will wrongly fire on a cook who literally spills beans. The field that has grown around this problem, usually abbreviated MWE research, tries to find such expressions in corpora, recognise them in running text, decide whether a given use is literal or figurative, and represent them in grammars, lexicons, parsers, translation systems and, more recently, neural language models.

The papers here run from a 1986 study of idioms in a rule-based translation system to shared tasks and model probes published in 2026. The topic overlaps closely with work on metaphor and figurative language, which shares many of its methods and datasets, and the article covers that overlap where the papers in this library do.

## What it is

### Idiomaticity in several dimensions

The most influential definition in the field, adopted by the survey of [Constant and colleagues](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) from Baldwin and Kim, describes MWEs as expressions that break into several lexemes and show *idiomaticity* at one or more levels. Idiomaticity can be lexical (a part does not occur on its own, as in *by dint of*), syntactic (the phrase breaks ordinary rules, as in *by and large*), semantic (the meaning is not built from the parts, as in *kick the bucket*), pragmatic (the phrase belongs to a fixed situation, as in *good morning*) or statistical (the combination is simply far more frequent than its alternatives, as in *strong tea* against *powerful tea*). Most expressions are idiomatic in more than one way, and the boundary with free combinations is graded rather than sharp.

### A classic typology

The paper that gave the field its slogan, [Multiword Expressions: A Pain in the Neck for NLP](https://md.research.elamur.ai/free/en/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/), proposed a classification that is still used. Lexicalised phrases split into *fixed* expressions that never vary, *semi-fixed* ones that inflect but otherwise stay put (non-decomposable idioms, many compounds, proper names), and *syntactically flexible* ones whose parts can be moved, passivised or modified (verb-particle constructions, decomposable idioms, light-verb constructions). *Institutionalised* phrases form a separate class: their meaning is compositional, but convention has frozen the wording. The authors argued that each class needs its own treatment in a grammar, and that neither the "words with spaces" approach nor full compositional analysis works across the board.

### Core notions

A few ideas recur through the whole literature:

- **Compositionality and decomposability.** An expression is compositional when its meaning follows from its parts. A decomposable idiom such as *pull strings* still maps its parts onto parts of its meaning (strings as connections), which is why it tolerates *strings were pulled*; a non-decomposable one such as *kick the bucket* does not.
- **Fixedness.** Idioms resist substituting one word for a near-synonym (*kick the pail* loses the idiom) and resist many syntactic changes. This resistance can be measured in corpora.
- **Discontinuity.** The parts of an expression need not be adjacent: *take this into account*, *the strings she pulled*.
- **Ambiguity between literal and idiomatic use.** Many expressions have a real literal reading, and some are used literally more often than not.
- **Types and tokens.** *Discovery* works at the level of expression types and produces lexicons; *identification* works on tokens and marks occurrences in text.

### Neighbours

Multiword expressions overlap with [lexical semantics and word embeddings](https://md.research.elamur.ai/free/en/topics/lexical-semantics/), since compositionality is a semantic question, and with [word sense disambiguation](https://md.research.elamur.ai/free/en/topics/word-sense-disambiguation/), since deciding whether *break the ice* is literal is in effect a sense decision. They are tied to syntax through [dependency parsing](https://md.research.elamur.ai/free/en/topics/dependency-parsing/), [treebanks](https://md.research.elamur.ai/free/en/topics/treebanks/) and [Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/), which must decide how to annotate an expression's internal structure. They differ from metaphor research in emphasis: a metaphor such as *grasp a theory* is often productive and novel, while an idiom is a stored, conventional item. The two fields nonetheless share methods, and many datasets mix idioms, metaphors and similes under the label of figurative language.

## History and milestones

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 1993–2017; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [1993 — Retrieving Collocations from Text: Xtract](https://md.research.elamur.ai/free/en/papers/smadja-1993-retrieving-collocations-text-xtract/) | 951 |
| [1993 — One Sense per Collocation](https://md.research.elamur.ai/free/en/papers/yarowsky-1993-one-sense-per-collocation/) | 440 |
| [1996 — Translating Collocations for Bilingual Lexicons: A Statistical Approach](https://md.research.elamur.ai/free/en/papers/smadja-1996-translating-collocations-bilingual-lexicons-statistical/) | 599 |
| [1999 — Automatic Identification of Non-compositional Phrases](https://md.research.elamur.ai/free/en/papers/lin-1999-automatic-identification-non-compositional-phrases/) | 288 |
| [2002 — Multiword Expressions: A Pain in the Neck for NLP](https://md.research.elamur.ai/free/en/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/) | 1,497 |
| [2003 — An Empirical Model of Multiword Expression Decomposability](https://md.research.elamur.ai/free/en/papers/baldwin-2003-empirical-model-multiword-expression-decomposability/) | 319 |
| [2009 — Unsupervised Type and Token Identification of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/fazly-2009-unsupervised-type-token-identification-idiomatic/) | 301 |
| [2011 — Literal and Metaphorical Sense Identification through Concrete and Abstract Context](https://md.research.elamur.ai/free/en/papers/turney-2011-literal-metaphorical-sense-identification-through/) | 394 |
| [2014 — Metaphor Detection with Cross-Lingual Model Transfer](https://md.research.elamur.ai/free/en/papers/tsvetkov-2014-metaphor-detection-cross-lingual-model/) | 369 |
| [2016 — Black Holes and White Rabbits: Metaphor Identification with Visual Features](https://md.research.elamur.ai/free/en/papers/shutova-2016-black-holes-white-rabbits-metaphor/) | 292 |
| [2017 — Identifying beneficial task relations for multi-task learning in deep neural networks](https://md.research.elamur.ai/free/en/papers/bingel-2017-identifying-beneficial-task-relations-multi/) | 305 |
| [2017 — Survey: Multiword Expression Processing: A Survey](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) | 296 |

### Rule-based beginnings

The earliest work treated idioms as a lexical problem inside large grammars. [Idioms in the Rosetta Machine Translation System](https://md.research.elamur.ai/free/en/papers/schenk-1986-idioms-rosetta-machine-translation-system/) described how the Philips Rosetta project, built on isomorphic grammars, could map a whole idiom structure onto one basic meaning, so that the ordinary grammar rules would apply to it and no special transfer was needed between languages.

### Collocation statistics

The arrival of large machine-readable corpora around 1990 moved attention to collocations. Following earlier association measures by Church and Hanks, Frank Smadja's [Retrieving Collocations from Text: Xtract](https://md.research.elamur.ai/free/en/papers/smadja-1993-retrieving-collocations-text-xtract/) combined corpus statistics with a parser to pull out word combinations, extend them to longer units and label them syntactically. Its sequel, [Translating Collocations for Bilingual Lexicons: A Statistical Approach](https://md.research.elamur.ai/free/en/papers/smadja-1996-translating-collocations-bilingual-lexicons-statistical/), used aligned parliamentary text to find French equivalents for English collocations.

### Non-compositionality and the "pain in the neck"

By the late 1990s the question shifted from frequency to meaning. [Automatic Identification of Non-compositional Phrases](https://md.research.elamur.ai/free/en/papers/lin-1999-automatic-identification-non-compositional-phrases/) by Dekang Lin tested whether a phrase behaves statistically like the variants obtained by swapping in similar words; an idiom stands out because its variants do not occur. In 2002 the Stanford LinGO group and colleagues published both the "pain in the neck" paper and a companion description of how the English Resource Grammar stores expressions in a database, [Multiword expressions: linguistic precision and reusability](https://md.research.elamur.ai/free/en/papers/lingbulk-2002-copestake-multiword-expressions-linguistic-precision-reusability/). The phrase stuck, and the ACL workshop series on multiword expressions began in 2003; its first edition included the [decomposability study](https://md.research.elamur.ai/free/en/papers/baldwin-2003-empirical-model-multiword-expression-decomposability/) of Baldwin and colleagues.

Papers on this topic by year 141 dated papers from 1986 to 2026; the busiest year is 2026 with 35, and 67% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 35 |
| 2025 | 10 |
| 2024 | 9 |
| 2023 | 15 |
| 2022 | 8 |
| 2021 | 4 |
| 2020 | 7 |
| 2019 | 2 |
| 2018 | 5 |
| 2017 | 5 |
| 2016 | 9 |
| 2015 | 1 |
| 2014 | 4 |
| 2013 | 2 |
| 2011 | 3 |
| 2010 | 3 |
| 2009 | 4 |
| 2007 | 1 |
| 2006 | 3 |
| 2004 | 1 |
| 2003 | 2 |
| 2002 | 2 |
| 1999 | 1 |
| 1996 | 1 |
| 1993 | 2 |
| 1991 | 1 |
| 1986 | 1 |

### Distributional models and idiom tokens

The 2000s brought vector-space models of meaning. Work by Baldwin, by [Katz and Giesbrecht](https://md.research.elamur.ai/free/en/papers/katz-2006-automatic-identification-non-compositional-multi/) and later by [Reddy, McCarthy and Manandhar](https://md.research.elamur.ai/free/en/papers/reddy-2011-empirical-study-compositionality-compound-nouns/) compared the vector of a whole expression with the vectors of its parts. In Toronto, Afsaneh Fazly, Paul Cook and Suzanne Stevenson turned the linguistic notion of fixedness into corpus measures and carried them from types to tokens, culminating in a 2009 *Computational Linguistics* article. In parallel, [Sporleder and Li](https://md.research.elamur.ai/free/en/papers/sporleder-2009-unsupervised-recognition-literal-non-literal/) used lexical cohesion to decide whether an idiom's words fit the surrounding text, a sign of literal use.

### Annotated corpora and shared tasks

The 2010s were the decade of data. Nathan Schneider and colleagues annotated every expression in a corpus of English web reviews, allowing gaps, and built supervised taggers on it; the [DiMSUM task](https://md.research.elamur.ai/free/en/papers/schneider-2016-semeval-2016-task-10-detecting/) at SemEval-2016 joined MWE identification with coarse semantic labels. The European PARSEME network (a COST Action) produced common guidelines for verbal expressions across many languages and ran [shared tasks in 2017](https://md.research.elamur.ai/free/en/papers/savary-2017-parseme-shared-task-automatic-identification/), [2018](https://md.research.elamur.ai/free/en/papers/ramisch-2018-edition-1-1-parseme-shared/) and [2020](https://md.research.elamur.ai/free/en/papers/ramisch-2020-edition-1-2-parseme-shared/), plus a [2017 survey](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) of the whole area. Large idiom corpora such as [MAGPIE](https://md.research.elamur.ai/free/en/papers/haagsma-2020-magpie-large-corpus-potentially-idiomatic/) followed.

### Neural models and language models

Since about 2020 most new papers ask how neural networks handle non-compositional meaning. Probing studies test whether contextual embeddings encode idiomaticity; machine translation studies show that Transformers translate idioms too literally; shared tasks at SemEval in [2022](https://md.research.elamur.ai/free/en/papers/madabushi-2022-semeval-2022-task-2-multilingual/) and [2025](https://md.research.elamur.ai/free/en/papers/pickard-2025-semeval-2025-task-1-admire/) test idiomaticity in text and then in images. The most recent wave evaluates large conversational models, asking whether they recall idioms from memory or reason about them, and the PARSEME effort has moved to a [2.0 corpus](https://md.research.elamur.ai/free/en/papers/savary-2026-parseme-2-0-multilingual-corpus/) and a [2.0 shared task](https://md.research.elamur.ai/free/en/papers/scholivet-2026-edition-2-0-parseme-shared/) that adds paraphrasing.

## Main problems and tasks

### Discovery

Discovery, also called extraction or acquisition, builds lists of candidate expressions from corpora. Classic pipelines generate n-grams or syntactic pairs, rank them by association measures such as pointwise mutual information, the Dice coefficient or log-likelihood, and filter them with part-of-speech patterns. Tools such as [mwetoolkit](https://md.research.elamur.ai/free/en/papers/aclbulk-2010-ramisch-multiword-expressions-wild-mwetoolkit-comes/) package these steps. Discovery also includes estimating properties of each type: how compositional it is, which syntactic forms it prefers, and what its canonical form looks like.

### Identification

Identification marks expression tokens in running text. It is harder than lookup because expressions inflect, take gaps and change order, and because the same word string may be literal. The PARSEME campaigns showed that the hardest case is not literal–idiomatic ambiguity but expressions that never appeared in the training data: systems that do well on seen items fall sharply on unseen ones.

### Literal versus figurative token classification

A narrower task asks, for a known potentially idiomatic expression, whether a given occurrence is meant figuratively. This is the setting of the Fazly–Cook–Stevenson work, of [sentence-embedding classifiers](https://md.research.elamur.ai/free/en/papers/salton-2016-idiom-token-classification-sentential-distributed/), of the first subtask of SemEval-2022, and of many recent evaluations of large language models.

### Compositionality prediction

Here the target is a graded score for an expression type: how far does *ivory tower* mean "a tower of ivory"? Datasets of human ratings for compound nouns, together with distributional models that compare whole and parts, define this line of work.

### Downstream use

Expressions matter for parsing, where they affect attachment and tokenisation; for [machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/), where literal renderings are among the most visible errors; for textual entailment and inference; for [sentiment and emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/), since figurative phrases carry evaluative charge; and for lexicography and language teaching. A recurring question, posed sharply in [Multi-word expressions in textual inference: Much ado about nothing?](https://md.research.elamur.ai/free/en/papers/aclbulk-2009-marneffe-multi-word-expressions-textual-inference/), is how much MWE handling actually changes end results.

## Methods and schools

### Lexicon and grammar engineering

The grammar-engineering school, rooted in HPSG and in rule-based translation, writes expressions into the lexicon with constraints on how they may vary. The English Resource Grammar's database, the Rosetta idiom treatment and the [expert lexicon for phrasal verbs](https://md.research.elamur.ai/free/en/papers/li-2003-expert-lexicon-approach-identifying-english/) of Li and colleagues are examples.

### Association measures and fixedness

The statistical school treats an expression as a combination that occurs more than chance predicts, or that resists variation more than a free phrase would. Lin's substitution test and the lexical and syntactic fixedness measures of [Fazly and Stevenson](https://md.research.elamur.ai/free/en/papers/fazly-2006-automatically-constructing-lexicon-verb-phrase/) belong here. These methods need no labelled data and transfer across languages, but they tend to mix idioms with merely frequent phrases.

### Distributional compositionality

Vector-space methods compare the observed vector of an expression with a vector composed from its parts. A large gap suggests non-compositional meaning. The approach began with LSA models and continued through additive and multiplicative composition functions, then through contextual embeddings.

### Sequence tagging

Once annotated corpora existed, identification could be framed as tagging. Schneider's group extended the familiar BIO scheme with lowercase tags for tokens inside a gap, allowing a first-order tagger to find discontinuous expressions efficiently. Later systems for PARSEME used conditional random fields, recurrent networks, graph-based parsers and, eventually, fine-tuned Transformers. [Tanner and Hoffman](https://md.research.elamur.ai/free/en/papers/tanner-2023-mwe-as-wsd-solving-multiword/) recast identification as sense disambiguation over a lexicon of candidates, using dictionary glosses to accept or reject each one.

### Neural representation and probing

With pretrained encoders, a large body of work asks where and whether idiomatic meaning is represented. Typical moves include replacing an expression by a synonym or by its parts and checking whether the embedding changes as it should, adding a single token for each expression to a model's vocabulary, reading off intermediate layers, and intervening on hidden states. The [survey by Miletić and Schulte im Walde](https://md.research.elamur.ai/free/en/papers/miletic-2024-semantics-multiword-expressions-transformer-based/) concludes that models capture this meaning only partly, lean on word overlap and memorised strings, and store what they know mostly in lower layers.

### Large language models

Recent work prompts generative models directly: to say whether a sentence contains an idiom, to explain it, to choose among paraphrases, to translate it, or to match it with an image. Results are mixed. Strong models do well on familiar English idioms, yet small open models tend to [see idioms where there are none](https://md.research.elamur.ai/free/en/papers/fornaciari-2024-hard-nut-crack-idiom-detection/), performance drops for lower-resource languages, and a recurring worry is that benchmark items have leaked into pretraining data.

## Datasets, resources and benchmarks

The field has a set of well-known shared resources, most of them described in papers in this library:

| Resource | What it contains |
| --- | --- |
| PARSEME corpora | Verbal MWEs annotated with common guidelines across many languages, used in the 2017, 2018 and 2020 shared tasks and extended in [PARSEME 2.0](https://md.research.elamur.ai/free/en/papers/savary-2026-parseme-2-0-multilingual-corpus/) |
| STREUSLE / CMWE | English web reviews with every MWE marked, including gappy ones, and later [supersense labels](https://md.research.elamur.ai/free/en/papers/missing-2015-schneider-corpus-model-integrating-multiword-expressions/) |
| DiMSUM | Reviews, tweets and TED talks with MWEs and supersenses, from SemEval-2016 |
| VNC-Tokens | English verb–noun combinations in British National Corpus sentences, labelled literal or idiomatic |
| MAGPIE | A large crowd-labelled English corpus of potentially idiomatic expressions drawn from the BNC |
| Noun Compound Senses and AStitchInLanguageModels | English and Portuguese compounds with paraphrases and fine-grained sense labels |
| ID10M | Silver-labelled idiom spans in ten languages, with a curated test set |
| FLUTE and IMPLI | Inference pairs that test whether models understand idioms, metaphors and similes |
| AdMIRe | Idiomatic compounds paired with images, in English and Brazilian Portuguese |

Lexical resources also matter: WordNet's multiword entries, Wiktionary idiom lists, idiom dictionaries, and specialised lexicons such as the [Hebrew verb–noun resource](https://md.research.elamur.ai/free/en/papers/lingbulk-2016-liebeskind-lexical-resource-hebrew-verb-noun/) of Chaya Liebeskind or the [English function-phrase dictionary](https://md.research.elamur.ai/free/en/papers/shigeto-2013-construction-english-mwe-dictionary-its/) built from Wiktionary by a Nara group. A recent [survey of idiom datasets](https://md.research.elamur.ai/free/en/papers/flor-2025-survey-idiom-datasets-psycholinguistic-computational/) maps what exists for both psycholinguistic and computational research.

## Landmark papers in this library

Citations of the papers on this topic, by when they were published Papers published in 2015–2019 gather 19% of the 11,692 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 45 papers, 99 citations |
| 2020–2024 | 43 papers, 1,710 citations |
| 2015–2019 | 22 papers, 2,241 citations |
| 2010–2014 | 12 papers, 1,996 citations |
| 2005–2009 | 8 papers, 869 citations |
| 2000–2004 | 5 papers, 2,216 citations |
| 1995–1999 | 2 papers, 887 citations |
| 1990–1994 | 3 papers, 1,660 citations |
| 1985–1989 | 1 papers, 14 citations |

### The 2002 typology paper

[Sag, Baldwin, Bond, Copestake and Flickinger](https://md.research.elamur.ai/free/en/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/) set out why expressions are frequent and troublesome, named idiomaticity and overgeneration as the two central problems, and proposed the fixed / semi-fixed / flexible / institutionalised typology described above. It remains the usual starting reference.

### Xtract (1993) and Champollion (1996)

[Xtract](https://md.research.elamur.ai/free/en/papers/smadja-1993-retrieving-collocations-text-xtract/) showed that a parser-based filter on top of association statistics yields far cleaner collocation lists than statistics alone; its lexicographic evaluation was run on stock-market news. [Champollion](https://md.research.elamur.ai/free/en/papers/smadja-1996-translating-collocations-bilingual-lexicons-statistical/), with Kathleen McKeown and Vasileios Hatzivassiloglou, grew French translation candidates word by word using the Dice coefficient over the Canadian Hansards, and handled collocations whose parts are separated, which earlier noun-phrase methods could not.

### Automatic Identification of Non-compositional Phrases (1999)

[Lin's paper](https://md.research.elamur.ai/free/en/papers/lin-1999-automatic-identification-non-compositional-phrases/) introduced the substitution test in its classic form: compare the mutual information of a phrase with that of variants built from thesaurus neighbours, and call it non-compositional when no variant comes statistically close. The idea of contrasting an expression with its lexical neighbours runs through much later work.

### An Empirical Model of Multiword Expression Decomposability (2003)

[Baldwin, Bannard, Tanaka and Widdows](https://md.research.elamur.ai/free/en/papers/baldwin-2003-empirical-model-multiword-expression-decomposability/) measured, in a latent semantic space, how similar compound nouns and verb-particle constructions are to their own parts. Raw similarity agreed poorly with WordNet, but higher-ranked items were more often WordNet hyponyms of their head, and the head was a far better predictor than the modifier. The paper made decomposability an empirical, measurable property.

### Fixedness and idiom tokens (2006–2009)

[Fazly and Stevenson](https://md.research.elamur.ai/free/en/papers/fazly-2006-automatically-constructing-lexicon-verb-phrase/) built lexical and syntactic fixedness measures for verb–noun pairs and a method for finding each idiom's canonical forms. [Cook, Fazly and Stevenson](https://md.research.elamur.ai/free/en/papers/cook-2007-pulling-their-weight-exploiting-syntactic/) then used those canonical forms to classify individual tokens, and the combined journal article, [Unsupervised Type and Token Identification of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/fazly-2009-unsupervised-type-token-identification-idiomatic/), showed that knowledge about the type can substitute for labelled examples of tokens. The VNC-Tokens data from this line became a standard test set.

### Multiword Expression Processing: A Survey (2017)

Written by members of the PARSEME network, [this survey](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) divided the field into discovery and identification and studied how each interacts with parsing and translation. Its notion of "orchestration", the question of when MWE processing should happen relative to the downstream task, gave the field a shared vocabulary.

### The PARSEME shared tasks (2017, 2018, 2020)

The [first edition](https://md.research.elamur.ai/free/en/papers/savary-2017-parseme-shared-task-automatic-identification/) created uniform guidelines for verbal expressions and annotated corpora in many languages; the [second](https://md.research.elamur.ai/free/en/papers/ramisch-2018-edition-1-1-parseme-shared/) revised the categories into a single decision tree and added measures by phenomenon, revealing a gap of roughly fifty F1 points between seen and unseen expressions; the [third](https://md.research.elamur.ai/free/en/papers/ramisch-2020-edition-1-2-parseme-shared/) was built around unseen items and found that the share of unseen expressions in a language predicts system quality much better than the size of its training set.

### Discriminative Lexical Semantic Segmentation with Gaps (2014)

[Schneider, Danchik, Dyer and Smith](https://md.research.elamur.ai/free/en/papers/schneider-2014-discriminative-lexical-semantic-segmentation-gaps/) built the first general-purpose supervised identifier for all kinds of English expressions, encoding gaps in an extended BIO scheme and distinguishing strong from weak expressions. The companion corpus, [Comprehensive Annotation of Multiword Expressions in a Social Web Corpus](https://md.research.elamur.ai/free/en/papers/lingbulk-2014-schneider-comprehensive-annotation-multiword-expressions-social/), became the core of later resources.

### MAGPIE (2020)

[Haagsma, Bos and Nissim](https://md.research.elamur.ai/free/en/papers/haagsma-2020-magpie-large-corpus-potentially-idiomatic/) produced an English corpus of potentially idiomatic expressions roughly ten times larger than earlier ones, with careful crowd annotation. It showed that most expressions lean strongly to one reading and that genre shifts the balance, and it became the input for many later neural studies.

### Can Transformer be Too Compositional? (2022)

[Dankers, Lucas and Titov](https://md.research.elamur.ai/free/en/papers/dankers-2022-can-transformer-be-too-compositional/) translated MAGPIE sentences with pretrained translation models and found that most figurative idioms were still rendered word for word. When a model did paraphrase an idiom, its encoder grouped the idiom's tokens more tightly and its decoder attended less to the source words. Removing figurativeness information from hidden states pushed outputs back to literal translations.

### SemEval-2022 Task 2 and AStitchInLanguageModels (2021–2022)

The Sheffield-led group of Harish Tayyar Madabushi, Aline Villavicencio and colleagues built a [dataset of noun compounds in context](https://md.research.elamur.ai/free/en/papers/madabushi-2021-astitchinlanguagemodels-dataset-methods-exploration-idiomaticity/) for English and Portuguese and turned it into a [SemEval task](https://md.research.elamur.ai/free/en/papers/madabushi-2022-semeval-2022-task-2-multilingual/) with Galician added as a low-resource test. Systems could detect idioms they had seen but generalised poorly to new ones, and sentence embeddings captured idiomatic meaning only weakly.

### Metaphor detection (2010–2016)

Several of the most cited papers on this topic concern metaphor. [Shutova, Sun and Korhonen](https://md.research.elamur.ai/free/en/papers/shutova-2010-metaphor-identification-verb-noun-clustering/) expanded a small seed set of verb metaphors by clustering nouns and verbs; [Turney and colleagues](https://md.research.elamur.ai/free/en/papers/turney-2011-literal-metaphorical-sense-identification-through/) used the abstractness of context words; [Tsvetkov and colleagues](https://md.research.elamur.ai/free/en/papers/tsvetkov-2014-metaphor-detection-cross-lingual-model/) built language-independent features and transferred an English model to other languages through dictionaries; and [Shutova, Kiela and Maillard](https://md.research.elamur.ai/free/en/papers/shutova-2016-black-holes-white-rabbits-metaphor/) added visual features from image search.

## People

Authors with the most papers on this topic Ekaterina Shutova leads with 9 papers; the 12 authors shown wrote 60 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Ekaterina Shutova](https://md.research.elamur.ai/free/en/authors/ekaterina-shutova/) | 9 papers · 1,466 citations |
| [Aline Villavicencio](https://md.research.elamur.ai/free/en/authors/aline-villavicencio/) | 9 papers · 470 citations |
| [Carlos Ramisch](https://md.research.elamur.ai/free/en/authors/carlos-ramisch/) | 7 papers · 749 citations |
| [Agata Savary](https://md.research.elamur.ai/free/en/authors/agata-savary/) | 7 papers · 408 citations |
| [Nathan Schneider](https://md.research.elamur.ai/free/en/authors/nathan-schneider/) | 5 papers · 490 citations |
| [Voula Giouli](https://md.research.elamur.ai/free/en/authors/voula-giouli/) | 4 papers · 375 citations |
| [Marco Idiart](https://md.research.elamur.ai/free/en/authors/marco-idiart/) | 4 papers · 227 citations |
| [Timothy Baldwin](https://md.research.elamur.ai/free/en/authors/timothy-baldwin/) | 3 papers · 1,896 citations |
| [Afsaneh Fazly](https://md.research.elamur.ai/free/en/authors/afsaneh-fazly/) | 3 papers · 474 citations |
| [Suzanne Stevenson](https://md.research.elamur.ai/free/en/authors/suzanne-stevenson/) | 3 papers · 474 citations |
| [Noah A. Smith](https://md.research.elamur.ai/free/en/authors/noah-a-smith/) | 3 papers · 390 citations |
| [Marie Candito](https://md.research.elamur.ai/free/en/authors/marie-candito/) | 3 papers · 372 citations |

[Aline Villavicencio](https://md.research.elamur.ai/free/en/authors/aline-villavicencio/), now at the University of Exeter and the University of Sheffield, is the most prolific author on this topic here. Her work spans from the 2002 grammar-engineering paper on the English Resource Grammar through mwetoolkit to the probing studies and SemEval tasks on idiomaticity in English, Portuguese and Galician.

[Carlos Ramisch](https://md.research.elamur.ai/free/en/authors/carlos-ramisch/), at CNRS and Aix-Marseille Université, built mwetoolkit, co-wrote the 2017 survey and led the 2018 and 2020 editions of the PARSEME shared task.

[Agata Savary](https://md.research.elamur.ai/free/en/authors/agata-savary/), at CNRS and Université Paris-Saclay, coordinated the PARSEME annotation effort, led the first shared task and the 2.0 corpus, and co-wrote the survey of how treebanks encode expressions.

[Timothy Baldwin](https://md.research.elamur.ai/free/en/authors/timothy-baldwin/), at the University of Melbourne and Mohamed bin Zayed University of Artificial Intelligence, co-authored the "pain in the neck" paper and the decomposability study, and with Su Nam Kim wrote the definition the field still cites.

[Ivan A. Sag](https://md.research.elamur.ai/free/en/authors/ivan-a-sag/), [Ann Copestake](https://md.research.elamur.ai/free/en/authors/ann-copestake/) (University of Cambridge), [Dan Flickinger](https://md.research.elamur.ai/free/en/authors/dan-flickinger/) and [Francis Bond](https://md.research.elamur.ai/free/en/authors/francis-bond/) (now at Palacký University Olomouc) form the grammar-engineering core behind the 2002 papers, which treated expressions inside a broad-coverage HPSG grammar.

[Frank Smadja](https://md.research.elamur.ai/free/en/authors/frank-smadja/) is the author of Xtract and, with [Kathleen McKeown](https://md.research.elamur.ai/free/en/authors/kathleen-mckeown/) and [Vasileios Hatzivassiloglou](https://md.research.elamur.ai/free/en/authors/vasileios-hatzivassiloglou/), of Champollion, the two foundational works on collocation extraction and translation.

[Afsaneh Fazly](https://md.research.elamur.ai/free/en/authors/afsaneh-fazly/), [Suzanne Stevenson](https://md.research.elamur.ai/free/en/authors/suzanne-stevenson/) (University of Toronto) and [Paul Cook](https://md.research.elamur.ai/free/en/authors/paul-cook/) developed the fixedness measures and unsupervised token classification for verb–noun idioms; Cook later also worked on token-level identification with multilingual pretrained models.

[Nathan Schneider](https://md.research.elamur.ai/free/en/authors/nathan-schneider/), now at Georgetown University, with [Noah A. Smith](https://md.research.elamur.ai/free/en/authors/noah-a-smith/) (Allen Institute), built the comprehensive English MWE annotation, the gap-aware tagger, the STREUSLE corpus and the DiMSUM shared task.

[Ekaterina Shutova](https://md.research.elamur.ai/free/en/authors/ekaterina-shutova/), at the University of Amsterdam, is the leading figure in computational metaphor here, from clustering-based detection and an annotated corpus of source–target mappings to multimodal detection and metaphor in political discourse.

[Peter Turney](https://md.research.elamur.ai/free/en/authors/peter-turney/) contributed the abstractness-based approach to literal and metaphorical senses and, with [Saif Mohammad](https://md.research.elamur.ai/free/en/authors/saif-mohammad/) and Shutova, an empirical study showing that metaphorical uses carry more emotion than literal ones.

[Marie Candito](https://md.research.elamur.ai/free/en/authors/marie-candito/), [Veronika Vincze](https://md.research.elamur.ai/free/en/authors/veronika-vincze/) and [Voula Giouli](https://md.research.elamur.ai/free/en/authors/voula-giouli/) are among the PARSEME organisers; Vincze also worked on detecting light-verb constructions across domains, and Giouli, based in Thessaloniki, contributed through the Greek PARSEME data up to the 2.0 corpus.

[Marco Idiart](https://md.research.elamur.ai/free/en/authors/marco-idiart/) (Universidade Federal do Rio Grande do Sul), [Carolina Scarton](https://md.research.elamur.ai/free/en/authors/carolina-scarton/) (Sheffield), [Marcos Garcia](https://md.research.elamur.ai/free/en/authors/marcos-garcia/) (Santiago de Compostela) and [Edward Gow-Smith](https://md.research.elamur.ai/free/en/authors/edward-gow-smith/) are Villavicencio's frequent collaborators on idiomaticity probes and shared tasks; Garcia has also worked on expressions in Galician translation.

[Thomas Pickard](https://md.research.elamur.ai/free/en/authors/thomas-pickard/) and [Maggie Mi](https://md.research.elamur.ai/free/en/authors/maggie-mi/) led the AdMIRe multimodal task, and Mi has recently revisited the hypothesis that decomposable idioms are more flexible.

[Yuji Matsumoto](https://md.research.elamur.ai/free/en/authors/yuji-matsumoto/), now at RIKEN and the Kyoto Institute of Technology, supervised work on an English dictionary of function-word expressions and on MWE-aware dependency corpora and parsing.

[Gülşen Eryiǧit](https://md.research.elamur.ai/free/en/authors/gulsen-eryi-it/) of Istanbul Technical University co-wrote the 2017 survey and has worked on multimodal idiomaticity benchmarks across languages. [Bruno Guillaume](https://md.research.elamur.ai/free/en/authors/bruno-guillaume/) (CNRS, Université de Lorraine) contributed to the PARSEME data and to the treatment of expressions in Surface-syntactic UD. [Chaya Liebeskind](https://md.research.elamur.ai/free/en/authors/chaya-liebeskind/) of the Jerusalem College of Technology works on Hebrew verb–noun expressions, and [Emmanuele Chersoni](https://md.research.elamur.ai/free/en/authors/emmanuele-chersoni/) of Hong Kong Polytechnic University on how humans and language models process idioms and light verbs.

## Institutions and places

Where the authors on this topic work now Centre National de la Recherche Scientifique is the current home of authors of 12 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [Centre National de la Recherche Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/) | 12 papers |
| [University of Amsterdam](https://md.research.elamur.ai/free/en/universities/university-of-amsterdam/) | 11 papers |
| [University of Exeter](https://md.research.elamur.ai/free/en/universities/university-of-exeter/) | 9 papers |
| [University of Sheffield](https://md.research.elamur.ai/free/en/universities/university-of-sheffield/) | 9 papers |
| [Aix-Marseille Université](https://md.research.elamur.ai/free/en/universities/aix-marseille-universite/) | 8 papers |
| [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) | 8 papers |
| [Laboratoire Interdisciplinaire des Sciences du Numérique](https://md.research.elamur.ai/free/en/labs/laboratoire-interdisciplinaire-des-sciences-du-numerique/) | 8 papers |
| [Laboratoire Parole et Langage](https://md.research.elamur.ai/free/en/labs/laboratoire-parole-et-langage/) | 8 papers |
| [Université Paris-Saclay](https://md.research.elamur.ai/free/en/universities/universite-paris-saclay/) | 8 papers |
| [Laboratoire d’Informatique et Systèmes](https://md.research.elamur.ai/free/en/labs/laboratoire-d-informatique-et-systemes/) | 7 papers |

The French national research centre, [Centre National de la Recherche Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/), is the leading current home of authors on this topic, largely because the PARSEME organisers are attached to it through laboratories such as the [Laboratoire d'Informatique et Systèmes](https://md.research.elamur.ai/free/en/labs/laboratoire-d-informatique-et-systemes/) and the [Laboratoire Parole et Langage](https://md.research.elamur.ai/free/en/labs/laboratoire-parole-et-langage/) at [Aix-Marseille Université](https://md.research.elamur.ai/free/en/universities/aix-marseille-universite/), the [Laboratoire Interdisciplinaire des Sciences du Numérique](https://md.research.elamur.ai/free/en/labs/laboratoire-interdisciplinaire-des-sciences-du-numerique/) at [Université Paris-Saclay](https://md.research.elamur.ai/free/en/universities/universite-paris-saclay/), and [Université de Lorraine](https://md.research.elamur.ai/free/en/universities/universite-de-lorraine/).

In Britain, the [University of Sheffield](https://md.research.elamur.ai/free/en/universities/university-of-sheffield/) and the [University of Exeter](https://md.research.elamur.ai/free/en/universities/university-of-exeter/) host the group behind the idiomaticity tasks of SemEval-2022 and SemEval-2025, and the [University of Cambridge](https://md.research.elamur.ai/free/en/universities/university-of-cambridge/) carries the grammar-engineering line. The [University of Amsterdam](https://md.research.elamur.ai/free/en/universities/university-of-amsterdam/) is the base of the metaphor work led by Shutova.

In the United States, [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) was where the comprehensive English annotation and the gap-aware tagger were built; its authors now include people at [Georgetown University](https://md.research.elamur.ai/free/en/universities/georgetown-university/), the [University of Colorado Boulder](https://md.research.elamur.ai/free/en/universities/university-of-colorado-boulder/), the [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/) and the [Allen Institute](https://md.research.elamur.ai/free/en/labs/allen-institute/). In Greece, the [Institute for Language and Speech Processing](https://md.research.elamur.ai/free/en/labs/institute-for-language-and-speech-processing/) of the [Athena Research Center](https://md.research.elamur.ai/free/en/labs/athena-research-and-innovation-center-in-information-communication-knowledge/) contributes through PARSEME, and [Hong Kong Polytechnic University](https://md.research.elamur.ai/free/en/universities/hong-kong-polytechnic-university/) through psycholinguistic studies.

Countries of the authors on this topic United States comes first: authors working there now wrote 41 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| United States | 41 papers |
| United Kingdom | 27 papers |
| Germany | 18 papers |
| France | 14 papers |
| Netherlands | 13 papers |
| Canada | 10 papers |
| China | 10 papers |
| Israel | 9 papers |
| Brazil | 8 papers |
| Romania | 6 papers |

Where the papers on this topic were published Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026) carries the most, 10 of 141.

**The numbers**

| Venue | Papers |
| --- | --- |
| Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026) | 10 papers |
| Proceedings of the Fifteenth Language Resources and Evaluation Conference | 10 papers |
| Computational Linguistics | 6 papers |
| Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16) | 4 papers |
| Transactions of the Association for Computational Linguistics | 4 papers |
| Proceedings of Learning Non-Literal Expressions with Small Data @ LREC 2026 | 3 papers |
| Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) | 3 papers |
| Glossa: a journal of general linguistics | 2 papers |
| Proceedings of the 12th Global Wordnet Conference | 2 papers |
| Proceedings of the 15th Joint Conference on Lexical and Computational Semantics (*SEM 2026) | 2 papers |

The main venue is the ACL workshop on multiword expressions, which has met regularly since 2003 and recently hosted the PARSEME 2.0 and AdMIRe 2 tasks. The LREC conference series, *Computational Linguistics*, *Transactions of the ACL* and the SemEval workshops carry much of the rest.

## Open questions

- **Unseen expressions.** Every major evaluation finds that systems handle expressions they have seen and fail on new ones. Whether better lexicons, unsupervised discovery or large models close this gap is still open.
- **What models actually know.** Probing results suggest that neural models store familiar idioms as memorised strings rather than understanding them, and studies of large models ask whether their apparent competence is [memorisation or reasoning](https://md.research.elamur.ai/free/en/papers/kim-2025-memorization-or-reasoning-exploring-idiom/). [Haviv and colleagues](https://md.research.elamur.ai/free/en/papers/haviv-2023-understanding-transformer-memorization-recall-through/) found a two-stage pattern in which early layers promote a memorised idiom ending and later layers boost its probability.
- **Literal bias in translation.** Neural translation still renders many idioms word for word. Retrieval and loss reweighting, as in [Liu, Chaudhary and Neubig](https://md.research.elamur.ai/free/en/papers/liu-2023-crossing-threshold-idiomatic-machine-translation/), help, and evaluation metrics such as those of [Baziotis and colleagues](https://md.research.elamur.ai/free/en/papers/baziotis-2023-automatic-evaluation-analysis-idioms-neural/) make the errors countable, but the problem is not solved.
- **Definitions and annotation.** Different resources draw the line between expression and free phrase differently, so a system can be "wrong" against one corpus and "right" against WordNet. Recent papers on [conventionalised but non-idiomatic expressions](https://md.research.elamur.ai/free/en/papers/katrapati-2025-non-idiomatic-conventionalised-expressions-new/) and on a [canonical form for flexible expressions](https://md.research.elamur.ai/free/en/papers/odijk-2024-canonical-form-flexible-multiword-expressions/) reopen these questions.
- **Languages beyond English.** Most data is still English. PARSEME covers many European languages, and newer work addresses [Chinese idiom paraphrasing](https://md.research.elamur.ai/free/en/papers/qiang-2023-chinese-idiom-paraphrasing/), [Bengali idioms](https://md.research.elamur.ai/free/en/papers/sakhawat-2026-when-words-don-t-mean/), [Marathi multiwords](https://md.research.elamur.ai/free/en/papers/padhye-2026-lock-stock-barrel-marathi-multiwords/), [Korean adpositional expressions](https://md.research.elamur.ai/free/en/papers/min-2026-curious-class-adpositional-multiword-expressions/) and [Hebrew idiom acquisition](https://md.research.elamur.ai/free/en/papers/lingbulk-2018-fadlon-acquisition-hebrew-idioms-stages-internal/), but coverage remains thin.
- **Human processing.** Psycholinguistics asks whether idioms and frequent phrases are retrieved whole. [Rambelli and colleagues](https://md.research.elamur.ai/free/en/papers/rambelli-2023-are-frequent-phrases-directly-retrieved/) found that readers process frequent compositional phrases about as fast as idioms, and that small language models mirror this better than large ones; [Are idioms surprising?](https://md.research.elamur.ai/free/en/papers/philipp-2023-are-idioms-surprising/) pursues similar questions with surprisal.
- **Meaning representation.** How to encode an idiom in a semantic graph is still debated, as in work on [UMR](https://md.research.elamur.ai/free/en/papers/bonn-2023-umr-annotation-multiword-expressions/) and [AMR](https://md.research.elamur.ai/free/en/papers/gareeva-2026-modelling-idiomatic-expressions-abstract-meaning/).

## Neighbouring topics

Neighbouring topics: papers shared Semitic languages shares the most: 16 of the 141 papers here (11%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) | 16 shared · 11% |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 14 shared · 10% |
| [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) | 9 shared · 6% |
| [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) | 8 shared · 6% |
| [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) | 6 shared · 4% |
| [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) | 6 shared · 4% |
| [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) | 6 shared · 4% |
| [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) | 5 shared · 4% |
| [Psycholinguistics & language acquisition](https://md.research.elamur.ai/free/en/topics/psycholinguistics/) | 4 shared · 3% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 3 shared · 2% |

The largest overlaps are with [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) and [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/), reflecting work on Hebrew verb–noun expressions, Hebrew idiom acquisition and flexibility, and idioms and metaphors in biblical and medieval Hebrew texts. The second-largest is [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/), where idioms serve as probes of memorisation and compositionality. Links to [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) and [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) come from multilingual idiom datasets and the literal-translation problem.

Metaphor and figurative-language work connects the topic with [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/), while inference benchmarks such as [FLUTE](https://md.research.elamur.ai/free/en/papers/chakrabarty-2022-flute-figurative-language-understanding-through/), [IMPLI](https://md.research.elamur.ai/free/en/papers/stowe-2022-impli-investigating-nli-models-performance/) and the narrative tasks of [It's not Rocket Science](https://md.research.elamur.ai/free/en/papers/chakrabarty-2022-it-s-not-rocket-science/) tie it to [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) and [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/). Image-based idiom tasks and meme analysis link it to [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/). Smaller overlaps exist with [Psycholinguistics & language acquisition](https://md.research.elamur.ai/free/en/topics/psycholinguistics/), [Dependency parsing](https://md.research.elamur.ai/free/en/topics/dependency-parsing/), [Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/), [Syntax & grammar formalisms](https://md.research.elamur.ai/free/en/topics/syntax-grammar/), [Lexical semantics & word embeddings](https://md.research.elamur.ai/free/en/topics/lexical-semantics/), [Historical linguistics & language change](https://md.research.elamur.ai/free/en/topics/language-change/) and [Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/).

The shelves this topic stands on Linguistics & languages: 117 of the 141 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) | 117 papers |
| [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) | 32 papers |
| [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) | 8 papers |
| [History & humanities](https://md.research.elamur.ai/free/en/shelves/history/) | 2 papers |
| [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) | 1 paper |

Almost every paper here also stands on the [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) shelf; smaller groups sit on [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) and [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/).

## Key works in this library

- [Idioms in the Rosetta Machine Translation System](https://md.research.elamur.ai/free/en/papers/schenk-1986-idioms-rosetta-machine-translation-system/) — idioms as single meanings in an isomorphic translation grammar.
- [Retrieving Collocations from Text: Xtract](https://md.research.elamur.ai/free/en/papers/smadja-1993-retrieving-collocations-text-xtract/) — statistics plus parsing for collocation extraction.
- [Translating Collocations for Bilingual Lexicons: A Statistical Approach](https://md.research.elamur.ai/free/en/papers/smadja-1996-translating-collocations-bilingual-lexicons-statistical/) — Champollion, collocation translation from aligned text.
- [Automatic Identification of Non-compositional Phrases](https://md.research.elamur.ai/free/en/papers/lin-1999-automatic-identification-non-compositional-phrases/) — the substitution test for idioms.
- [Multiword Expressions: A Pain in the Neck for NLP](https://md.research.elamur.ai/free/en/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/) — the standard typology and statement of the problem.
- [Multiword expressions: linguistic precision and reusability](https://md.research.elamur.ai/free/en/papers/lingbulk-2002-copestake-multiword-expressions-linguistic-precision-reusability/) — storing expressions in a broad HPSG grammar.
- [An Empirical Model of Multiword Expression Decomposability](https://md.research.elamur.ai/free/en/papers/baldwin-2003-empirical-model-multiword-expression-decomposability/) — decomposability measured in vector space.
- [An Expert Lexicon Approach to Identifying English Phrasal Verbs](https://md.research.elamur.ai/free/en/papers/li-2003-expert-lexicon-approach-identifying-english/) — constrained lexicon entries for particle verbs.
- [Automatic Identification of Non-Compositional Multi-Word Expressions using Latent Semantic Analysis](https://md.research.elamur.ai/free/en/papers/katz-2006-automatic-identification-non-compositional-multi/) — composed versus observed vectors for German.
- [Unsupervised Type and Token Identification of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/fazly-2009-unsupervised-type-token-identification-idiomatic/) — fixedness measures from types to tokens.
- [Unsupervised Recognition of Literal and Non-Literal Use of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/sporleder-2009-unsupervised-recognition-literal-non-literal/) — lexical cohesion as a literalness cue.
- [An Empirical Study on Compositionality in Compound Nouns](https://md.research.elamur.ai/free/en/papers/reddy-2011-empirical-study-compositionality-compound-nouns/) — human ratings for compounds and their parts.
- [Discriminative Lexical Semantic Segmentation with Gaps: Running the MWE Gamut](https://md.research.elamur.ai/free/en/papers/schneider-2014-discriminative-lexical-semantic-segmentation-gaps/) — general supervised identification with gaps.
- [SemEval-2016 Task 10: Detecting Minimal Semantic Units and their Meanings (DiMSUM)](https://md.research.elamur.ai/free/en/papers/schneider-2016-semeval-2016-task-10-detecting/) — joint expressions and supersenses.
- [MWEs in Treebanks: From Survey to Guidelines](https://md.research.elamur.ai/free/en/papers/bank-2016-rosen-mwes-treebanks-survey-guidelines/) — how treebanks encode expressions, with principles.
- [Survey: Multiword Expression Processing: A Survey](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) — discovery, identification and orchestration.
- [The PARSEME Shared Task on Automatic Identification of Verbal Multiword Expressions](https://md.research.elamur.ai/free/en/papers/savary-2017-parseme-shared-task-automatic-identification/) — the first multilingual verbal benchmark.
- [Edition 1.2 of the PARSEME Shared Task on Semi-supervised Identification of Verbal Multiword Expressions](https://md.research.elamur.ai/free/en/papers/ramisch-2020-edition-1-2-parseme-shared/) — focus on unseen expressions.
- [MAGPIE: A Large Corpus of Potentially Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/haagsma-2020-magpie-large-corpus-potentially-idiomatic/) — the large English idiom corpus.
- [Probing for idiomaticity in vector space models](https://md.research.elamur.ai/free/en/papers/garcia-2021-probing-idiomaticity-vector-space-models/) — substitution probes for compounds in English and Portuguese.
- [Can Transformer be Too Compositional? Analysing Idiom Processing in Neural Machine Translation](https://md.research.elamur.ai/free/en/papers/dankers-2022-can-transformer-be-too-compositional/) — why translation models go literal.
- [ID10M: Idiom Identification in 10 Languages](https://md.research.elamur.ai/free/en/papers/tedeschi-2022-id10m-idiom-identification-10-languages/) — multilingual idiom tagging from silver data.
- [SemEval-2022 Task 2: Multilingual Idiomaticity Detection and Sentence Embedding](https://md.research.elamur.ai/free/en/papers/madabushi-2022-semeval-2022-task-2-multilingual/) — idiomaticity in classification and embeddings.
- [Semantics of Multiword Expressions in Transformer-Based Models: A Survey](https://md.research.elamur.ai/free/en/papers/miletic-2024-semantics-multiword-expressions-transformer-based/) — what Transformers capture of expression meaning.
- [SemEval-2025 Task 1: AdMIRe - Advancing Multimodal Idiomaticity Representation](https://md.research.elamur.ai/free/en/papers/pickard-2025-semeval-2025-task-1-admire/) — matching idioms to images.
- [PARSEME 2.0 Multilingual Corpus of Multiword Expressions](https://md.research.elamur.ai/free/en/papers/savary-2026-parseme-2-0-multilingual-corpus/) — the next generation of PARSEME data.

## All 141 papers on this topic

**1** [2](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/2/) … [6](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/6/) [next →](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/2/)

- [Multiword Expressions: A Pain in the Neck for NLP](https://md.research.elamur.ai/free/en/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/) — Ivan A. Sag, Timothy Baldwin, Francis Bond, et al. · 2002 · 1 497 citations
- [Retrieving Collocations from Text: Xtract](https://md.research.elamur.ai/free/en/papers/smadja-1993-retrieving-collocations-text-xtract/) — Frank Smadja · 1993 · 951 citations
- [Translating Collocations for Bilingual Lexicons: A Statistical Approach](https://md.research.elamur.ai/free/en/papers/smadja-1996-translating-collocations-bilingual-lexicons-statistical/) — Frank Smadja, Kathleen R. McKeown, Vasileios Hatzivassiloglou · 1996 · 599 citations
- [One Sense per Collocation](https://md.research.elamur.ai/free/en/papers/yarowsky-1993-one-sense-per-collocation/) — David Yarowsky · 1993 · 440 citations
- [Literal and Metaphorical Sense Identification through Concrete and Abstract Context](https://md.research.elamur.ai/free/en/papers/turney-2011-literal-metaphorical-sense-identification-through/) — Peter Turney, Yair Neuman, Dan Assaf, et al. · 2011 · 394 citations
- [Metaphor Detection with Cross-Lingual Model Transfer](https://md.research.elamur.ai/free/en/papers/tsvetkov-2014-metaphor-detection-cross-lingual-model/) — Yulia Tsvetkov, Leonid Boytsov, Anatole Gershman, et al. · 2014 · 369 citations
- [An Empirical Model of Multiword Expression Decomposability](https://md.research.elamur.ai/free/en/papers/baldwin-2003-empirical-model-multiword-expression-decomposability/) — Timothy Baldwin, Colin Bannard, Takaaki Tanaka, et al. · 2003 · 319 citations
- [Identifying beneficial task relations for multi-task learning in deep neural networks](https://md.research.elamur.ai/free/en/papers/bingel-2017-identifying-beneficial-task-relations-multi/) — Joachim Bingel, Anders Søgaard · 2017 · 305 citations
- [Unsupervised Type and Token Identification of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/fazly-2009-unsupervised-type-token-identification-idiomatic/) — Afsaneh Fazly, Paul Cook, Suzanne Stevenson · 2009 · 301 citations
- [Survey: Multiword Expression Processing: A Survey](https://md.research.elamur.ai/free/en/papers/constant-2017-survey-multiword-expression-processing-survey/) — Mathieu Constant, Gülşen Eryiǧit, Johanna Monti, et al. · 2017 · 296 citations
- [Black Holes and White Rabbits: Metaphor Identification with Visual Features](https://md.research.elamur.ai/free/en/papers/shutova-2016-black-holes-white-rabbits-metaphor/) — Ekaterina Shutova, Douwe Kiela, Jean Maillard · 2016 · 292 citations
- [Automatic Identification of Non-compositional Phrases](https://md.research.elamur.ai/free/en/papers/lin-1999-automatic-identification-non-compositional-phrases/) — Dekang Lin · 1999 · 288 citations
- [Metaphor as a Medium for Emotion: An Empirical Study](https://md.research.elamur.ai/free/en/papers/mohammad-2016-metaphor-as-medium-emotion-empirical/) — Saif Mohammad, Ekaterina Shutova, Peter Turney · 2016 · 285 citations
- [CorMet: A Computational, Corpus-Based Conventional Metaphor Extraction System](https://md.research.elamur.ai/free/en/papers/mason-2004-cormet-computational-corpus-based-conventional/) — Zachary J. Mason · 2004 · 270 citations
- [met*: A Method for Discriminating Metonymy and Metaphor by Computer](https://md.research.elamur.ai/free/en/papers/fass-1991-met-method-discriminating-metonymy-metaphor/) — Dan Fass · 1991 · 269 citations
- [SemEval-2020 Task 8: Memotion Analysis- the Visuo-Lingual Metaphor!](https://md.research.elamur.ai/free/en/papers/sharma-2020-semeval-2020-task-8-memotion/) — Chhavi Sharma, Deepesh Bhageria, William Scott, et al. · 2020 · 260 citations
- [Metaphor Identification Using Verb and Noun Clustering](https://md.research.elamur.ai/free/en/papers/shutova-2010-metaphor-identification-verb-noun-clustering/) — Ekaterina Shutova, Lin Sun, Anna Korhonen · 2010 · 250 citations
- [An Empirical Study on Compositionality in Compound Nouns](https://md.research.elamur.ai/free/en/papers/reddy-2011-empirical-study-compositionality-compound-nouns/) — Siva Reddy, Diana McCarthy, Suresh Manandhar · 2011 · 233 citations
- [The PARSEME Shared Task on Automatic Identification of Verbal Multiword Expressions](https://md.research.elamur.ai/free/en/papers/savary-2017-parseme-shared-task-automatic-identification/) — Agata Savary, Carlos Ramisch, Silvio Cordeiro, et al. · 2017 · 205 citations
- [Neural Metaphor Detection in Context](https://md.research.elamur.ai/free/en/papers/gao-2018-neural-metaphor-detection-context/) — Ge Gao, Eunsol Choi, Yejin Choi, et al. · 2018 · 198 citations
- [Statistical Metaphor Processing](https://md.research.elamur.ai/free/en/papers/shutova-2013-statistical-metaphor-processing/) — Ekaterina Shutova, Simone Teufel, Anna Korhonen · 2013 · 196 citations
- [Unsupervised Recognition of Literal and Non-Literal Use of Idiomatic Expressions](https://md.research.elamur.ai/free/en/papers/sporleder-2009-unsupervised-recognition-literal-non-literal/) — Caroline Sporleder, Linlin Li · 2009 · 173 citations
- [Automatic Identification of Non-Compositional Multi-Word Expressions using Latent Semantic Analysis](https://md.research.elamur.ai/free/en/papers/katz-2006-automatic-identification-non-compositional-multi/) — Graham Katz, Eugenie Giesbrecht · 2006 · 172 citations
- [A Report on the 2018 VUA Metaphor Detection Shared Task](https://md.research.elamur.ai/free/en/papers/leong-2018-report-2018-vua-metaphor-detection/) — Chee Wee (Ben) Leong, Beata Beigman Klebanov, Ekaterina Shutova · 2018 · 167 citations
- [Metaphor Corpus Annotated for Source - Target Domain Mappings](https://md.research.elamur.ai/free/en/papers/shutova-2010-metaphor-corpus-annotated-source-target/) — Ekaterina Shutova, Simone Teufel · 2010 · 164 citations

**1** [2](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/2/) … [6](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/6/) [next →](https://md.research.elamur.ai/free/en/topics/multiword-expressions/page/2/)

- [← previous topic Historical linguistics & language change](https://md.research.elamur.ai/free/en/topics/language-change/)
- [next topic → Implicature](https://md.research.elamur.ai/free/en/topics/implicature/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)