💡 Cognitive linguistics
Cognitive linguistics is an approach to language that treats grammar and meaning as products of general human cognition, including categorisation, perception, bodily experience and imagination. It does not treat language as an autonomous formal module, and it identifies meaning closely with conceptualisation.
What it is
Cognitive linguistics is a family of theories, not a single model. They share a few commitments. The first is that linguistic descriptions should fit what is known about the mind and brain. The second is that language uses the same mechanisms as other cognitive abilities, such as attention, memory, analogy and perspective-taking. The third is that grammar and lexicon form a continuum of form–meaning pairings learned from use.
Meaning on this view is encyclopedic and perspectival. A word gives access to a structured body of knowledge, and the same situation can be construed in different ways, for example as an active or a passive scene, or as a process or a thing. Categories are graded rather than defined by necessary and sufficient features. Abstract thought is largely organised through mappings from concrete, embodied experience.
The approach sits between linguistics, psychology and philosophy. Since the late 1980s it has also fed computational work, above all on metaphor and metonymy (Fass 1991).
The key idea
The best-known mechanism is the conceptual metaphor. Here a target domain is understood by systematically projecting structure onto it from a source domain. Computational work has operationalised proxies for parts of this idea. One example is a word's abstractness, which serves as a proxy for whether a context belongs to an abstract target domain:
M : S → T (TARGET IS SOURCE, e.g. ARGUMENT IS WAR)
abs(w) = Σᵢ sim(w, aᵢ) − Σⱼ sim(w, cⱼ), rescaled to [0, 1]
In the first line, S is a source domain such as WAR, JOURNEY or CONTAINER, and T is a target domain such as ARGUMENT, LIFE or IDEAS. M is a set of correspondences between their elements, for instance opponents mapping to arguers and attacks mapping to criticisms.
In the second line, w is a word and sim is cosine similarity in a distributional vector space. Each aᵢ is one of a set of abstract paradigm words and each cⱼ is one of a set of concrete ones. Turney and colleagues used 20 of each, chosen with ratings from the MRC Psycholinguistic Database. Their scorer correlated at 0.81 with human abstractness ratings on a held-out half of 4,295 rated words (Turney et al. 2011). The working hypothesis is that a word surrounded by abstract context is more likely to be used metaphorically, because metaphor moves knowledge from concrete sources into abstract targets. Scores of the surrounding words therefore serve as the classifier's features (Turney et al. 2011).
Metonymy is the companion mechanism. A metonymy is a "stands for" relation within one domain, whereas a metaphor is a "seen as" relation across domains. Fass made this contrast operational with his met* method, implemented in the program meta5. In met*, a semantic relation is tested in turn as literal, metonymic, metaphorical and finally anomalous (Fass 1991).
History and founding papers
The movement took shape in the late 1970s and 1980s as a reaction against formal, modular grammar. Its core theories and figures are:
| Theory | Central idea | Founding figures |
|---|---|---|
| Conceptual metaphor | Abstract domains structured by embodied source domains | Lakoff and Johnson 1980 |
| Prototypes and categorisation | Graded categories with central and peripheral members | Lakoff 1987, building on the psychologist Rosch (1975) as an influence |
| Frame semantics | Words evoke structured scenes of background knowledge | Fillmore 1982 |
| Cognitive grammar | Grammar as symbolic pairings and construal | Langacker 1987 |
| Construction grammar | Constructions carry meaning independently of words | Fillmore, Kay and O'Connor 1988; Goldberg 1995 |
| Mental spaces and blending | Online integration of partial conceptual packets | Fauconnier 1985; Fauconnier and Turner 2002 |
| Embodiment and image schemas | Meaning grounded in recurrent bodily patterns | Johnson 1987; Talmy 1988 |
Lakoff and colleagues catalogued source–target mappings in the Master Metaphor List. Later annotation schemes drew on this list. In one verb-metaphor corpus, the source categories covered all metaphorical expressions in a held-out set and the target categories covered 76% (Shutova & Teufel 2010). That scheme built on the Pragglejaz Group's MIP and its notion of a more basic, bodily and older sense (Shutova & Teufel 2010).
Computational modelling began with knowledge-based systems. Fass's meta5 program was presented as the first working system to recognise both metonymy and metaphor, though only on a small set of sentences (Fass 1991). Lapata and Lascarides gave a probabilistic, corpus-learned account of logical metonymy, as in "enjoy the book" read as enjoying reading it. They ranked implicit events in a way that tracked human paraphrase judgments, and they contrasted their method with Pustejovsky's qualia roles (Lapata & Lascarides 2003). CorMet discovered conventional mappings from web corpora and WordNet. It noticed, for example, that pour takes funds in finance but liquids in the lab, and it recovered part of the Master Metaphor List (Mason 2004).
Statistical methods followed. Shutova, Sun and Korhonen started from 62 seed metaphors and expanded them through verb and noun clustering, reaching 0.79 precision (Shutova et al. 2010). A later system added interpretation by literal paraphrase (Shutova et al. 2013). Abstractness-based classifiers generalised to unseen verbs (Turney et al. 2011). Tsvetkov and colleagues transferred an English model to Spanish, Farsi and Russian through bilingual dictionaries. They read this as support for metaphor being conceptual rather than lexical (Tsvetkov et al. 2014).
How it is used today
In linguistics, cognitive frameworks remain standard tools for semantic description, including description of ancient languages. In Neo-Assyrian letters, compound prepositions built from body-part nouns mostly express directed motion or extensions of it. About 11% of them are non-trivial conceptual metaphors (Ong & Gordin 2024). Translators of the Hebrew Bible into Afrikaans used conceptual metaphor and blending to argue for keeping images such as "God is a sun" rather than turning them into similes (Van der Merwe & Hendriks 2006).
In NLP, metaphor detection is now a benchmarked task built mainly on MIPVU-annotated corpora. The first shared task on the VU Amsterdam corpus had a best All-POS F1 of 0.651. Verbs were the easiest part of speech and adverbs the hardest (Leong et al. 2018). The 2020 edition added learner essays annotated for argumentation-relevant metaphors. There the best All-POS F1 rose to 0.769 on VUA and 0.715 on TOEFL (Leong et al. 2020).
Contextual embeddings drove much of this progress. A GloVe-plus-ELMo BiLSTM reached 72.6 F1 on VUA sequence labelling (Gao et al. 2018). Perceptual grounding, a computational echo of embodiment, also helps: combining visual and textual embeddings beat either modality alone (Shutova et al. 2016).
Applications extend in several directions:
- Emotion: metaphorical verb uses were rated more emotional than literal ones (Mohammad et al. 2016).
- Political framing: metaphor detection as an auxiliary task improved models of political perspective, affiliation and framing (Cabot et al. 2020).
- Sentiment: an Arabic sentiment tool degraded on metaphorical sentences (Alsiyat & Piao 2020).
- Multimodal material: internet memes, which pair an image with a caption, were loosely framed as visuo-lingual metaphor in a benchmark that actually targets sentiment, humour and offensiveness (Sharma et al. 2020).
- New languages and periods: German religious forums (Reimann & Scheffler 2026), classical Japanese (82.18 F1 under 5-fold cross-validation) (Zhu et al. 2026) and an Arabic adaptation of MIPVU (Marathe et al. 2026).
Large language models have become both objects and tools of study. On the MUNCH benchmark, models often confused source and target domains (Tong et al. 2024). In another study, models had to pick one of 9 intention categories to explain why a metaphor was used. GPT-4 reached about 45% accuracy and Llama2-70B-Chat peaked at 39.00%; all models beat a random baseline but stayed weak (Michelli et al. 2026). Metaphorical mapping can also serve as an attack. In evaluations on AdvBench and HarmBench, benign analogues of harmful requests achieved over 92% attack success on GPT-4o within 3 retries (Yan et al. 2025).
Open questions
-
Conceptual or lexical? The cross-lingual transfer result described above points towards conceptual mappings (Tsvetkov et al. 2014). Transfer in German, however, worked mainly where both languages share a mapping (Reimann & Scheffler 2026). How universal and how culture-specific mappings are remains unresolved.
-
Reliable identification. Many neural errors are borderline cases that reflect subjective annotation (Gao et al. 2018). Arabic annotation showed low Fleiss' κ, partly because over 90% of tokens were jointly labelled non-metaphorical, which inflates chance agreement (Marathe et al. 2026). Performance also varies more by genre than by part of speech (Leong et al. 2018).
-
Novelty and creativity. LM surprisal correlates only moderately with human judgments of metaphor novelty, with a best r of .49 (Momen et al. 2026). LLM-written metaphors are hard for people to tell apart from human ones but easy for classifiers to detect (Regneri et al. 2026).
-
Interpretation, not only detection. Metaphor caused mistranslations in 27 of 62 test sentences in one machine-translation pilot (Shutova et al. 2013). Fluency metrics can penalise the literal paraphrases that understanding requires (Sanchez-Montero et al. 2026).
-
Where meaning comes from. Emotional force seems to emerge from combining source and target rather than being inherited from the source (Mohammad et al. 2016). This is arguably the kind of pattern blending theory would predict, but computational models rarely implement blending itself.
Beyond metaphor, the field still debates how strong the embodiment claim is, how prototypes relate to formal semantics, and whether construction grammar can be made fully explicit and learnable.
Key papers
Related concepts
- Construction grammar — Construction grammar is a family of linguistic theories holding that knowledge of a language consists of constructions. Constructions are…
- Psycholinguistics — Psycholinguistics is the study of the mental processes by which people produce, comprehend and acquire language. It combines experimental…
- 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…
- Implicature — Implicature is the part of what a speaker communicates that goes beyond the literal content of the words. Hearers infer it by assuming the…
More papers
Topics: 🧒 Psycholinguistics & language acquisition · 📖 Lexical semantics & word embeddings
Shelves: 🗣️ Linguistics & languages
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/cognitive-linguistics/