Verbal Multiword Expressions

• VMWEs are multi-token units with idiosyncratic syntactic and semantic properties, combining non-compositional meaning with inflection and discontinuity.
• They are classified into subtypes such as verbal idioms, verb–particle, light-verb, and support-verb constructions, each varying in fixedness and linguistic behavior.
• Research leverages sequence tagging, syntax-aware models, and transformer-based approaches to address challenges in detection, interpretation, and translation of VMWEs.

Verbal multiword expressions (VMWEs) are multi-token units whose syntactic behavior and semantics are idiosyncratic relative to regular compositional syntax, and whose head or core is verbal. They include constructions such as verbal idioms, verb–particle constructions, light-verb constructions, support-verb constructions, inherently reflexive verbs, and multi-verb constructions. Across theoretical linguistics and NLP, VMWEs are treated as a major subclass of MWEs because they combine lexicalized meaning with inflection, argument structure, discontinuity, and cross-linguistic variation in especially difficult ways; the same string may be literal in one sentence and idiomatic in another, and the relevant signal is often distributed across sentence context rather than localized in a contiguous surface span (Taslimipoor et al., 2018, Miletić et al., 2024).

1. Taxonomic scope and linguistic categories

In the PARSEME framework, VMWEs are organized into several recurrent subtypes: light-verb constructions, verbal idioms, inherently reflexive verbs, verb–particle constructions, multi-verb constructions, and inherently adpositional verbs. The categories reported in the multilingual shared-task literature are LVC.full, LVC.cause, VID, IRV, VPC.full, VPC.semi, MVC, and IAV. These categories differ in fixedness, semantic idiosyncrasy, and syntactic variability: some light-verb constructions are relatively patterned, whereas verbal idioms are typically highly idiosyncratic (Taslimipoor et al., 2018).

Before the table, one distinction is essential. Not all VMWE typologies cut the space in the same way. PARSEME uses fine-grained verbal categories; CoAM uses a coarse PoS-based type VERB for any MWE headed by a verb, and separates some subject-bearing phatic or proverbial expressions into CLAUSE rather than VERB. This means that the same English item might be a VID or VPC in PARSEME-oriented work but only VERB in CoAM (Ide et al., 2024).

Category	Core characterization	Typical behavior
LVC.full / LVC.cause	light-verb construction	support verb plus nominal predicate
VID	verbal idiom	strongly non-compositional
VPC.full / VPC.semi	verb–particle construction	particle shift, frequent discontinuity
IRV	inherently reflexive verb	reflexive-marked unit
MVC	multi-verb construction	verbal chain functioning as one predicate
IAV	inherently adpositional verb	verb requires specific adposition

Support-verb constructions, also called light-verb constructions, occupy a particularly important position within this taxonomy. In French Lexicon-Grammar work, an SVC is a predicate realized through a support verb and a predicative noun, with the noun carrying the core lexical meaning and the verb contributing inflectional, aspectual, or stylistic information. This perspective emphasizes that some VMWEs are better understood as predicate realizations than as fixed strings, because the predicative noun may preserve argument structure even without an overt support verb (0902.3072).

A more formal theoretical treatment appears in radically lexicalized CCG analyses of VMWEs. There, verb–particle constructions, idiomatically combining phrases, and phrasal idioms are separated by how their components contribute to predicate structure. Phrasal idioms such as kick the bucket are modeled with singleton syntactic types, whereas idiomatically combining phrases such as spill the beans use head-marked polyvalent categories. The central claim is paracompositionality: some parts of a VMWE contribute not to ordinary argument structure but to contingency, such as manner, aspect, or event-type information (Bozsahin et al., 2018).

2. Compositionality, idiomaticity, and syntactic variability

A standard starting point is the literal–idiomatic contrast. In idiomaticity detection work, MWEs are defined as expressions of at least two words that are syntactically or semantically specific. Literal usage is compositional or approximately compositional; idiomatic usage is conventionalized and non-compositional. For VMWEs this distinction is especially unstable, because verb sense, argument structure, and surrounding collocations are all context-sensitive. The same verbal expression can therefore alternate between literal and idiomatic readings across sentences (Chu et al., 2022).

Discontinuity is a central structural property. A continuous VMWE occupies one contiguous span, but a discontinuous VMWE may have its components separated by modifiers, objects, passivization, coordination, or language-specific word-order effects. Examples discussed in the literature include English make … an effort, take … a look, and picked one up, as well as French, German, and Persian cases where syntactic dependency, rather than adjacency, links the components. This is one reason why local n-gram methods and strictly sequential assumptions are systematically fragile for VMWEs (Rohanian et al., 2019).

Inflectional and syntactic variation compound the problem. VMWE instances may differ by tense, agreement, passivization, clitic placement, scrambling, or insertion of modifiers. This is particularly visible in VMWE-oriented reinterpretations of contextual idiomaticity detection: exact surface matching “without deformation” is serviceable for broad MWE benchmarks, but brittle for VMWEs because inflected and discontinuous realizations are common. Lemma-aware matching, component-wise marking, and syntactic anchoring are therefore plausible necessities for VMWE-specific systems (Chu et al., 2022).

Support-verb constructions show that VMWE variability is not only a matter of discontinuity. In a large French corpus study, predicative nouns were identified 95,430 times, but support-verb constructions overlapping those noun occurrences were identified only 3,349 times. The experimental proportion of predicative nouns accompanied by a support verb was therefore approximately 4%, rising to 8% after correction via the paper’s precision–recall adjustment formula $n' = np/r$. This means that overt support-verb realizations were a minority, and that recognition strategies restricted to verb+noun co-occurrence would miss most realizations of this subtype (0902.3072).

Idiomaticity is also graded rather than absolute. In an expert-annotated Russian dataset of 286 MWEs evaluated on 16 criteria, there were no absolutely idiomatic expressions; lexical criteria were the most influential, whereas grammatical criteria were “bound to certain conditions.” The study further observed that the presence of a verb usually makes expressions less fixed in Russian, which suggests that VMWEs may often combine high lexicalized meaning with comparatively broader morphosyntactic latitude (Mikhalkova et al., 19 May 2026).

3. Corpora, annotation frameworks, and evaluation protocols

VMWE research is anchored in several distinct but partly overlapping resource traditions: multilingual identification corpora, idiomaticity-in-context datasets, parallel corpora for translation, and recent all-type MWE benchmarks.

Resource	Scope	Notable property
PARSEME 1.1 / 1.2	multilingual VMWE identification	fine-grained VMWE categories
SemEval-2022 Task 2	idiomatic vs literal usage	given MWE mention in context
VNC-Tokens	English VNC idiomaticity	token-level literal/idiomatic labels
AlphaMWE	parallel vMWE corpus	manual bilingual alignment
CoAM	English all-type MWE identification	VERB type, discontinuous/overlapping spans

PARSEME Shared Task 1.1 provided annotated corpora for 20 languages, with native speakers annotating both continuous and discontinuous VMWEs and marking category labels such as VID or LVC.full. To adapt these data to sequence tagging, SHOMA converts the original format into an IOB-like scheme in which the first token of a VMWE receives B-Category, subsequent components receive I-Category, and non-VMWE tokens receive O. Rare overlaps are encoded as semicolon-concatenated labels treated as atomic classes (Taslimipoor et al., 2018). A later PARSEME edition, used by MTLB-STRUCT, evaluates 14 languages and reports both Global and Unseen expression-level F1, making generalization to unseen VMWE types a first-class evaluation target (Taslimipoor et al., 2020).

SemEval-2022 Task 2 addresses a different problem: idiomaticity detection for a provided MWE mention in context. In Subtask A, the languages are English, Portuguese, and Galician, with reported test set sizes of 916, 713, and 713 respectively. The label is binary—idiomatic or literal—and the data format includes the target MWE type and the sentence containing its realization, so the task is usage classification rather than span discovery. Zero-shot and one-shot conditions make the benchmark useful for studying cross-lingual and low-shot contextual semantics (Chu et al., 2022).

English idiomaticity resources often center on verb–noun combinations. VNC-Tokens contains 2,984 sentences and 56 expressions, each sentence containing exactly one VNC; one probing study uses a binary subset of 1,205 sentences, with 749 Idiomatic and 456 Literal instances. The same work explicitly redesigns train–test splits so that test VNCs are unseen during training, reducing lexical memorization and forcing abstraction over idiomatic usage (Klubička et al., 2023).

Recent resource construction has moved in two directions. AlphaMWE provides 750 English sentences with vMWE tags aligned to manually post-edited Chinese, German, and Polish translations, retaining verbal-head anchoring and cross-lingual span alignment (Han et al., 2020). CoAM, by contrast, covers all MWE types in English and supports discontinuous and overlapping spans; it contains 1,301 sentences, 30,231 tokens, a test set of 385 MWEs, and an unseen rate of 64.2% by lemma multiset. In that test set, the VERB type accounts for 154 instances, or 40.0% (Ide et al., 2024).

Evaluation protocols vary by task, but two formulas dominate. For identification, expression-level precision, recall, and F1 are standard:

$$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$

For idiomaticity detection, macro-F1 is often preferred:

$F1_{\text{macro}}=\frac{1}{K}\sum_{k=1}^{K} F1_k,$

with $K=2$ for idiomatic and literal classes (Chu et al., 2022).

4. Modeling approaches

Sequence tagging remains the dominant formulation for VMWE identification. SHOMA uses a language-independent architecture that combines fixed 300-dimensional fastText embeddings, seven binary word-shape features, one-hot POS tags, two convolutional layers with kernel sizes 2 and 3, a BiLSTM with 300 hidden units, and an optional linear-chain CRF. The CRF did not consistently improve MWE-based F1 across languages, but it accelerated convergence in some settings. A simple post-processing rule—filtering I-Category predictions without a preceding B-Category—improved strict MWE-based precision (Taslimipoor et al., 2018).

Discontinuity-focused work makes the representation problem explicit. A VMWE identifier based on ELMo embeddings combines two parallel branches before a BiLSTM tagger: a Graph Convolutional Network over dependency adjacency matrices and a multi-head self-attention branch preceded by multi-channel CNNs. The model uses an IOBG tagging scheme, where G marks tokens in the gap of a discontinuous VMWE. Highway layers integrate syntax-aware and content-aware signals before token classification (Rohanian et al., 2019).

Joint learning with syntax is another recurrent design. MTLB-STRUCT fine-tunes multilingual BERT while sharing encoder layers between VMWE tagging and dependency parsing. VMWE prediction is a token-level softmax layer; the auxiliary parsing task uses bilinear arc scoring and a TreeCRF over unlabeled dependency trees. The multitask loss is

$$\mathcal{L}=\lambda_{vmwe}\mathcal{L}_{vmwe}+\lambda_{dep}\mathcal{L}_{dep},$$

with $\lambda_{vmwe}=1$ and $\lambda_{dep}=\alpha$, where $\alpha$ is selected per language from $\{1/300,1/700\}$ (Taslimipoor et al., 2020).

For idiomaticity classification when the candidate expression is already known, contextual sentence encoders dominate. In the SemEval-2022 system from HIT, only the target sentence is used; the MWE surface form is marked with tokenizer [SEP] tokens when it appears exactly; the encoder output is summarized by the [CLS] representation; and classification is performed with a fully connected layer with dropout and softmax:

$p(y\mid x)=\mathrm{softmax}(Wh+b).$

Training uses cross-entropy, and robustness is strengthened with R-Drop and adversarial training. The same paper argues that for VMWEs, exact surface marking should be replaced by component-aware and lemma-aware strategies that can tolerate discontinuity and inflection (Chu et al., 2022).

A very different line of work focuses on previously seen VMWEs and hand-engineered variability profiles. A French system identifies occurrences of known VMWE types by first extracting candidates and then classifying them with a linear SVM using only six features. The optimal feature set is dominated by insertion patterns and lexical identity: REL_insertSeq, ABS_insertSeq, REL_0to5insertions, ABS_lemma_VERB, REL_lemma_VERB, and ABS_LemmaSet (Pasquer et al., 2020).

5. Empirical results and what they show

The multilingual PARSEME results established two durable findings: neural sequence models can generalize beyond memorized VMWE types, but discontinuity remains a persistent error source. SHOMA achieved an official macro-average MWE-based F1 of 58.09 across languages, with token-based macro F1 63.40, and was especially strong on unseen data entries. On five languages used for seen/unseen analysis, it consistently outperformed a transition-based parser on unseen types; for example, EN unseen was 16.45 F1 versus 13.30, and FR unseen was 36.32 versus 16.94 (Taslimipoor et al., 2018).

Explicitly targeting discontinuity yields measurable gains. On PARSEME test data for English, German, French, and Persian, the hybrid GCN+attention model outperformed both strong baselines and prior systems in most settings. In English, for instance, the hybrid reached 41.91 MWE-based F1 on all MWEs and 22.73 on the discontinuous subset, compared with SHOMA’s 26.42 and 1.90 respectively. In French, the hybrid reached 70.97 overall and 63.90 on discontinuous MWEs, exceeding SHOMA’s 62.00 and 51.43 (Rohanian et al., 2019).

Joint syntactic supervision pushes multilingual performance further. MTLB-STRUCT ranked first in the PARSEME 2020 open track with overall macro Global F1 $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$0 and Unseen F1 $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$1, and achieved the best Unseen F1 on 8 of 14 languages. The strongest predictor of Global F1 was the proportion of seen types in training, with reported correlation $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$2, but auxiliary dependency parsing still improved unseen detection in several languages, including Turkish, Swedish, Portuguese, and Greek (Taslimipoor et al., 2020).

Contextual idiomaticity classification reaches higher absolute scores because the task is easier once the candidate expression is given. In the SemEval-2022 setup, XLM-R large with R-Drop reached dev macro-F1 of approximately 80.34 in zero-shot and approximately 92.91 in one-shot; an ensemble of 13 models obtained test F1 $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$3 in zero-shot and $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$4 in one-shot, ranking first in one-shot. The same study reports that XLM-R consistently outperformed mBERT, larger models continued to help, and naive token-level span pooling underperformed the sentence-level [CLS] representation (Chu et al., 2022).

Across transformers more broadly, representation studies are mixed rather than uniformly optimistic. A survey of transformer-based MWE semantics concludes that models capture MWE and VMWE meaning inconsistently, rely heavily on surface patterns and memorized information, and localize much of the useful signal in early or mid layers rather than in the final layer. Seen–unseen–cross-lingual degradation is systematic: one mBERT idiomaticity result on English PIEs falls from 83.8% accuracy on seen items to 74.3% on unseen English and 72.4% on Russian; in Slovene PIEs, mBERT yields lower seen F1 than a trilingual model but higher unseen F1 and better transfer to Croatian and Polish (Miletić et al., 2024).

Probing confirms that contextual encoders contain idiomaticity information, but not in a particularly robust way. On English VNCs, mean-pooled BERT sentence embeddings reached AUC 0.8411 on the fixed unseen-VNC split and 0.8524 on resampled splits, while mean-pooled GloVe reached 0.7485 and 0.7717. In the same experiments, ablating embedding dimensions drove performance to chance, whereas ablating the norm left performance almost unchanged, suggesting that the recoverable idiomaticity signal resides in direction rather than magnitude (Klubička et al., 2023).

Recent general-purpose LLMs have not removed the difficulty. In CoAM, the best fine-tuned model, Qwen-72B, achieved overall micro F1 $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$5, with VERB recall $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$6, and the paper explicitly notes that Verb MWEs were easier than Noun MWEs across approaches. Yet recall on unseen items and non-WordNet items remained substantially lower (Ide et al., 2024). A multilingual and code-switched evaluation of LLMs on idiomaticity detection likewise found that GPT-4 did not outperform a fine-tuned xlm-roBERTa-base baseline overall: in zero-shot detection, the baseline achieved macro-F1 0.6540, whereas GPT-4 achieved $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$7; in semantic interpretation, the baseline achieved 0.7712 versus GPT-4’s $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$8. Performance degraded further on synthetic MWEs and code-switched inputs, indicating substantial reliance on familiarity and memorization (Leon et al., 10 Apr 2025).

6. Downstream effects, limitations, and research directions

The importance of VMWEs is not confined to identification benchmarks. In machine translation, they remain a measurable bottleneck. A 2025 study covering English to Czech, German, Russian, Chinese, Japanese, Spanish, and Turkish reports that VMWEs consistently degrade translation quality, with severity ordered by non-compositionality: verbal idioms are hardest, verb–particle constructions intermediate, and light-verb constructions least problematic. The paper gives concrete failure cases such as He’s dead to the world translated into Chinese as “他已经死了,” and it took two to tango rendered literally with tango in several target languages (Liu et al., 24 Aug 2025).

That finding aligns with AlphaMWE’s corpus-construction experience. During manual post-editing and bilingual annotation of 750 English vMWE sentences translated into Chinese, German, and Polish, accurate translation of MWEs proved sufficiently difficult that the project introduced an error taxonomy covering common sense, super sense, abstract phrase, idiom, metaphor, and several forms of ambiguity. The resulting alignments show that a source VMWE may map to a target idiom, a compositional paraphrase, or even a single lexical verb, so category-preserving transfer cannot be assumed (Han et al., 2020).

Mitigation strategies now increasingly combine detection with meaning-preserving rewriting. In the MT study, an LLM-based paraphrasing pipeline replaced VMWE spans with literal counterparts before translation and improved quality most for verbal idioms and VPCs. Representative MetricX-24 gains include English→German with Opus on VIDs, where Ori was 6.87 and $$\mathrm{Precision}=\frac{TP}{TP+FP}, \quad \mathrm{Recall}=\frac{TP}{TP+FN}, \quad F1=\frac{2PR}{P+R}.$$9, and English→Czech with Opus on VIDs, where Ori was 13.41 and $F1_{\text{macro}}=\frac{1}{K}\sum_{k=1}^{K} F1_k,$0. The same study reports that LVC paraphrasing was much less consistently beneficial, consistent with their higher compositionality (Liu et al., 24 Aug 2025).

Several limitations recur across the literature. VMWE phenomena are heterogeneous; data are sparse for rare idioms and low-resource languages; discontinuity and morphology undermine exact string matching; syntactic auxiliaries depend on parser quality; and evaluation protocols remain difficult to compare across corpora and languages. The transformer survey therefore argues for more directly comparable setups, while idiomaticity-detection work points toward component-aware span marking, adapters with syntactic signals, and contrastive objectives tailored to idiomatic vs literal pairs rather than generic sentence-level contrastive learning (Miletić et al., 2024, Chu et al., 2022).

A coherent research agenda is now visible. One direction is structural: integrate POS tags, dependency arcs, or multitask parsing so that non-adjacent verbal components can be modeled compositionally and discontinuously. Another is lexical-semantic: inject idiom dictionaries, paraphrase supervision, or sense inventories, especially for rare and unseen VMWEs. A third is evaluative: standardize multilingual benchmarks that jointly stress identification, subtype classification, idiomaticity judgment, and generalization to unseen or code-switched settings. The recent evidence that large models still underperform specialized multilingual encoders on nuanced MWE tasks suggests that scale alone has not solved VMWE processing (Taslimipoor et al., 2020, Leon et al., 10 Apr 2025).

Taken together, the literature presents VMWEs as a convergence point of lexicalized meaning, structured syntax, and contextual interpretation. They are not merely anomalous phrases to be memorized; they are recurring loci where grammar, semantics, and cross-lingual variation interact, and where progress depends on combining linguistic structure with contextual representation rather than replacing one with the other.