🏗️ Construction grammar
Construction grammar is a family of linguistic theories holding that knowledge of a language consists of constructions. Constructions are conventional pairings of form and meaning at every size and level of abstraction, from words and idioms to abstract clause patterns, and they are linked in a structured network.
What it is
Construction grammar (CxG) rejects a strict split between a lexicon of meaningful words and a separate, meaning-free rule system. It describes linguistic knowledge as one network of form–meaning pairs, which can be concrete or highly schematic. Because there is no dividing line between lexicon and grammar, the same machinery handles transparent, compositional expressions and idiosyncratic, non-compositional ones (Anonymous 2025). The central claim is that structures at every level, not only words, contribute meaning (Anonymous 2023).
The units range along a continuum of schematicity:
- A fully fixed idiom such as "kick the bucket".
- Partly open patterns such as "the X-er, the Y-er" or "X let alone Y".
- Fully abstract argument-structure patterns such as the ditransitive.
This graded range has been used as a design dimension for test sets. One corpus groups constructions by how schematic they are, in order to probe how far learners and models abstract (Bonial & Madabushi 2024).
The key idea
A construction is a pair of a form and a meaning. An utterance is licensed, that is, sanctioned as well formed, when compatible constructions combine.
C = ⟨F, M⟩
Ditransitive: [ Subj V Obj1 Obj2 ] ↔ 'X causes Y to receive Z' (central sense)
Way: [ Subj V POSS way PATH ] ↔ 'X moves along PATH by (or while) V-ing'
The symbols mean the following:
- C is a construction, F its form side and M its meaning side.
- The arrow ↔ marks the conventional link between the two sides.
- Subj, Obj1 and Obj2 are syntactic slots. V is the verb slot, POSS a possessive determiner coreferential with the subject, and PATH a directional phrase.
- X, Y and Z are the semantic roles that the slots fill.
The main consequence is that meaning can come from the pattern itself rather than from the verb.
The ditransitive is polysemous. Its central sense is actual transfer, as in "She gave him a cake", while verbs of creation such as "bake" license a related sense of intended transfer. In "She baked him a cake", the verb "bake" does not express transfer, yet the sentence means that she intended him to receive the cake. The ditransitive construction contributes this transfer meaning. The verb is fused with it, which means the verb's participant roles are unified with the construction's argument roles.
In "He elbowed his way through the crowd", "elbow" is not a motion verb, and "way" is not a real object. The way-construction supplies the motion along a path, while the verb supplies the manner or means.
Caused-motion patterns such as "sneeze the napkin off the table" have been formalized in Embodied Construction Grammar (Dodge & Petruck 2014). A related "V NP1 into NP2" pattern has been used to resolve prepositional-phrase attachment through semantic features (Chen et al. 2011).
Constructions are organized by inheritance links. A specific pattern inherits properties from more general ones. When a word's own meaning clashes with a construction's meaning, the construction can coerce a new reading of the word, that is, shift its interpretation to fit the pattern.
History and founding papers
The idea grew out of work on idioms and frame semantics at Berkeley, associated with Fillmore, Kay and O'Connor (1988). It drew on Lakoff (1987) and on Langacker's Cognitive Grammar (1987). Goldberg (1995) made argument-structure constructions the centre of the theory, and Croft (2001) proposed Radical Construction Grammar.
On the usage-based side, Tomasello (2003) and Bybee (2010) are standard references for the view that constructions are learned from experience. Among the computational branches, Steels developed grounded, constructivist models of how agents build construction grammars (Steels 2004). This line led to Fluid Construction Grammar (FCG), which is introduced briefly in (Steels & Beule 2006).
| Variant | Typical association | Characteristic emphasis |
|---|---|---|
| Berkeley Construction Grammar | Fillmore, Kay (1988 onward) | Unification-based feature structures, idioms, frames |
| Goldberg's (Cognitive) Construction Grammar | Goldberg (1995, 2006) | Argument-structure constructions, inheritance, usage-based learning |
| Radical Construction Grammar | Croft (2001) | Constructions as primitives, categories defined per construction and per language |
| Cognitive Grammar (closely related framework) | Langacker (1987) | Symbolic units of semantic and phonological poles, conceptualization |
| Fluid Construction Grammar | Steels and colleagues (2000s onward) | Executable bidirectional processing, language evolution |
FCG grew into an extensive software platform for both parsing and production. Its developers presented it in system demonstrations (Trijp et al. 2012; Wellens et al. 2013). A volume on design patterns for Fluid Construction Grammar, edited by Luc Steels, which sets out the formalism's design-pattern methodology, was the subject of a book review (Schneider & Tsarfaty 2013). A notable application used FCG experiments to argue that the German case system is less arbitrary than it appears (van Trijp 2012). A later overview summarized the formalism's status and prospects (Beuls & Eecke 2023).
Constructional analysis has also reached beyond argument structure. One example is antonymic coordination in French, analysed within a framework presented as a multidimensional construction grammar (Guénot 2008).
How it is used today
Usage-based learning is a major strand. In this view, constructions emerge from repeated exposure, and frequency and analogy shape their productivity. Computational work operationalizes this in several ways:
- Comparing learned grammars by complexity and descriptive adequacy (Dunn 2018).
- Showing that grammars learned from different registers converge as exposure grows (Dunn & Madabushi 2021).
- Combining distributional semantics with CxG as a unified usage-based model (Rambelli et al. 2019).
FCG researchers learn large-scale grammars from semantically annotated corpora (Eecke & Beuls 2026). They also model how constructions can arise from sensory experience (Ekila et al. 2025) and combine symbolic constructions with embeddings (Verheyen et al. 2025).
A continuous alternative is Stretched Tree Metric Grammars, one subtype of the broader class of metric grammars. This approach treats grammaticality as graded "harmony" in a metric space. In a toy English simulation, the novel denominal verb's share of the forms in the training data was raised from 0% to 8%. Each condition was tested on 10 items, and results were averaged across 5 runs. The added exposure lifted the verb's harmony to a level comparable with attested forms, while cross-category novelties stayed below it (Tabor & Lee 2026).
Annotation is another active area:
- A treebank scheme for English argument structure constructions (Sung & Kyle 2024).
- A Python tool for measuring how texts use those constructions (Sung & Kyle 2025).
- Annotation of verb-argument structure guided by usage-based analogy (Hsiao & Michaelis 2025).
Earlier work proposed enriching semantic representations with constructional information to make them less domain-dependent (Hwang et al. 2010). Recent frameworks extend computational CxG to signed languages (Vos et al. 2025).
Constructions in language models are now a prominent use. CxGBERT brought constructional information into BERT (Madabushi et al. 2020). Weissweiler and colleagues argued that CxG offers insight into neural models that other frameworks miss (Weissweiler et al. 2023). Later work used constructional evidence to argue that large models solve tasks by extrapolating from context (Madabushi & Bonial 2025).
A dedicated workshop series supports this community. The first edition met at Georgetown University in March 2023, as part of GURT/SyntaxFest (Anonymous 2023). The second met in Düsseldorf on 24 September 2025, co-located with the 16th IWCS conference. Of its 35 submissions, 17 archival papers were presented in person, along with 5 online papers and 9 non-archival posters (Anonymous 2025).
Open questions
Several questions remain unsettled:
- Evaluation is unresolved. How should one judge whether a computational construction grammar is correct or useful (Marques & Beuls 2016)? Automatic identification of schematic clausal constructions has produced instructive negative results, for example in Brazilian Portuguese (Lorenzi et al. 2023).
- For language models, it is still debated whether success on constructional probes reflects genuine abstraction or surface cues. Datasets graded by schematicity are one attempt to separate the two (Bonial & Madabushi 2024).
- There is a theoretical tension between discrete construction inventories and gradient, continuously changing grammar. The metric-grammar authors list several caveats about their toy simulation. Introducing the novel verb appeared to reduce the harmony of existing constructions, the model has not been scaled up, and its robustness has not been tested (Tabor & Lee 2026).
- Further questions concern how far constructional analyses carry across languages and modalities, and how symbolic and neural representations of constructions can be integrated (Verheyen et al. 2025; Vos et al. 2025).
Key papers
Related concepts
- Cognitive linguistics — Cognitive linguistics is an approach to language that treats grammar and meaning as products of general human cognition, including…
- Context-free grammar — A context-free grammar is a formal system of rewriting rules in which each rule replaces a single nonterminal symbol with a string of…
- Child-directed speech — Child-directed speech is the speech register adults and older children use when talking to young children. It typically has higher and more…
- Lexical semantics — Lexical semantics is the study of word meaning. It asks what words and multiword expressions denote, how a word's senses are organized, how…
More papers
Topics: 📐 Syntax & grammar formalisms · 🧒 Psycholinguistics & language acquisition
Shelves: 🗣️ Linguistics & languages
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/construction-grammar/