Natural language processing systems traditionally rely on compositionality—the principle that the meaning of an expression is built predictably from the literal meanings of its parts. However, figurative language intentionally breaks these rules. When a speaker says "the corporate ladder is slippery" or "he devoured the book," they are applying concrete properties to abstract concepts.
To detect these shifts automatically, computational models look for selectional preference violations. A selectional preference (or restriction) is a semantic constraint that a word places on its neighbors. For instance, the verb devour typically prefers an animate subject and a physical, consumable object. When these constraints are breached—such as an inanimate object performing an action reserved for living beings—it serves as a robust statistical indicator of non-literal language.
As Ekaterina Shutova notes in Models of Metaphor in NLP, traditional semantic models often fail when encountering these combinations because they treat them as logical errors. By reframing these clashes as intentional figurative choices, systems can identify the need for metaphorical processing when standard literal interpretations fail to satisfy linguistic constraints (Yorick Wilks et al. in Your Metaphor or Mine: Belief Ascription and Metaphor Interpretation).
Quantifying Violations with Statistical Strength
How do we measure a semantic violation without manually hand-coding every rule? The answer lies in calculating selectional preference strength (SPS), a metric that quantifies how strictly a word restricts its arguments based on corpus statistics.
The authors of Statistical Metaphor Processing demonstrate that measuring SPS allows a system to treat low-probability argument pairings as potential figurative language. If a verb has high preference strength (meaning it is highly selective about its arguments), a violation is highly informative. Conversely, verbs with weak constraints are less useful filters, a property leveraged by Ekaterina Shutova et al. in Metaphor Identification Using Verb and Noun Clustering to isolate highly predictive metaphorical verbs.
To scale this across diverse vocabularies, systems map words to coarse-grained semantic categories or ontologies rather than relying on exact word matches. For example, Yulia Tsvetkov et al. in Metaphor Detection with Cross-Lingual Model Transfer use WordNet "supersenses" to detect when a verb's expected semantic class for an object is replaced by an incompatible one, such as pairing a consumption verb with a non-food substance.
Similarly, Cheng Yang et al. in Can ChatGPT's Performance be Improved on Verb Metaphor Detection Tasks? Bootstrapping and Combining Tacit Knowledge extract syntactic arguments and compare their topical categories to pre-defined literal domains to systematically map these boundaries.
Literal Usage: [The chef] -> devours -> [the steak]
(Animate) (Food)
| |
Match Match
Metaphorical Shift: [The student] -> devours -> [the book]
(Animate) (Abstract/Object)
| |
Match VIOLATION (Metaphor Signal)
Measuring Tension in Representation Space
With the advent of deep learning, the concept of a selectional preference violation shifted from discrete category clashes to vector space incongruity. Instead of checking if a noun belongs to a "food" category, neural models measure the semantic tension between a word's contextualized representation and its isolated, literal definition.
This is often implemented via late-interaction or Siamese architectures. For instance, Minjin Choi 0001 et al. introduce MelBERT in MelBERT: Metaphor Detection via Contextualized Late Interaction using Metaphorical Identification Theories, which compares the global embedding of a sentence with the contextualized embedding of a target word. If the target word's meaning shifts drastically away from its literal baseline to fit the sentence, a violation is detected.
This principle is mirrored in the work of Shenglong Zhang and Ying Liu in Adversarial Multi-task Learning for End-to-end Metaphor Detection, which explicitly models the discrepancy between a word's context and its most frequent, concrete meaning. To handle the inherent uncertainty of literal word boundaries, Wenbo Qiao et al. in A Quantum-Inspired Matching Network with Linguistic Theories for Metaphor Detection combine selectional preference violation detection with quantum probability principles to capture contextual incongruity dynamically.
Domain Mismatches and Concreteness Shifts
At its core, a selectional preference violation in a metaphor reflects a conceptual domain mismatch: applying the rules of a concrete source domain to an abstract target domain.
The authors of Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining formalize this by measuring the distance between the domain of a target argument and the domain of a verb's typical literal arguments (e.g., physical consumption vs. information processing). This perspective is supported by Lisa Gandy et al. in Automatic Identification of Conceptual Metaphors With Limited Knowledge, who identify metaphors by flagging instances where concrete descriptors are applied to abstract entities.
This shift can also be tracked at the discourse level. Hyeju Jang et al. show in Metaphor Detection in Discourse that representing words through standardized concreteness scores allows models to flag semantic clashes when highly concrete words are dropped into abstract thematic environments, creating a break in lexical coherence. For comparative figurative language like similes, Vlad Niculae demonstrates in Comparison pattern matching and creative simile recognition that measuring the semantic distance between primary entities serves as a proxy for identifying figurative language, where low similarity suggests a creative relationship between distant concepts.
Nuances and Exceptions
While selectional preference violations are highly effective for identifying explicit metaphors, natural language presents several edge cases where literal constraints are bent but not broken metaphorically.
One such phenomenon is argument coercion, where a verb forces an object into a different semantic category to maintain a literal or metonymic reading (e.g., "I'm reading Shakespeare," where the author stands for the book). As noted by Kirk Roberts et al. in Unsupervised Learning of Selectional Restrictions and Detection of Argument Coercions, this flexibility complicates automated analysis.
Furthermore, Katja Markert et al. demonstrate in Metonymy Resolution as a Classification Task that figurative language frequently occurs in entirely neutral contexts that do not violate any literal logical constraints. This means rule-based violation detectors must be augmented with statistical likelihood models to capture subtle shifts where a literal interpretation remains technically possible but statistically improbable.
Key Foundational Work
A System of Programs for Computer-Aided Induction: A Summary (1977) — contributed 2 concept elements to this direction.
A System of Programs for Computer-Aided Induction: A Summary (1977) — contributed 1 concept elements to this direction.
Metonymy and Metaphor: What’s the Difference? (1988) — contributed 1 concept elements to this direction.