Overcoming Categorical Brittleness with Flexible Semantic Constraints

23 concepts22 papers2 proposals

When early natural language processing systems encountered the sentence "The car drank gasoline," they typically ground to a halt. Under traditional frameworks, the verb drink carries a strict selection restriction: its subject must be animate, and its object must be a liquid. Because a car is an inanimate artifact, a rigid rule-based parser flags this sentence as a semantic violation and rejects it.

Yet, humans understand the sentence effortlessly. We do not treat semantic constraints as rigid, binary gates that block comprehension; instead, we treat them as flexible preferences. The research direction explored across these papers addresses this exact challenge: replacing categorical semantic brittleness—where language is forced into rigid, all-or-nothing ontological boxes—with flexible, continuous, and adaptive semantic constraints.

From Binary Logic to Preference Semantics

In the early days of artificial intelligence, the dominant paradigm treated language comprehension as a form of mathematical theorem-proving. If a sentence could be parsed into a valid logical formula that satisfied all ontological constraints, it was "understood."

Yorick Wilks challenged this notion in Understanding Without Proofs, introducing a paradigm known as Preference Semantics. He argued that human-like understanding is not about calculating a mathematical proof, but about finding the most plausible interpretation given the context. Instead of rejecting a sentence for violating a semantic norm, a system should prefer interpretations that maximize semantic density and common-sense patterns. If no perfectly literal interpretation is available, the system does not crash; it accepts the best available match.

Rigid Selection Restrictions:
[Car: Inanimate] -> Verb: Drink [Requires Animate] -> ERROR: Reject Input

Preference Semantics:
[Car: Inanimate] -> Verb: Drink [Prefers Animate]  -> Relax Constraint -> Match "Consume Fuel"

This shift from binary gates to continuous evaluation was formalized by G., who proposed representing semantic constraints as continuous, fuzzy functions rather than all-or-nothing properties. By linking the degree of an interpretation's sensibleness to a spectrum of plausibility, this approach allowed systems to model real-world language use more accurately. As Pinkal et al. later detailed, binary selection restrictions fundamentally fail to account for the soft constraints and degrees of preference that characterize everyday communication.

Dynamic Type Shifting and Coercion

If semantic constraints are flexible, how do systems systematically resolve violations without manual intervention for every possible metaphor? The answer lies in mechanisms that dynamically alter or "coerce" semantic types based on context.

One such mechanism is sort coercion, introduced by Hays et al.. When a system encounters a low-level metaphor—such as applying spatial length to time (e.g., "a long afternoon"), it dynamically re-casts the ontological types. By allowing the system to view a temporal event through a spatial lens, sort coercion bridges the gap between rigid database schemas and flexible human speech.

This flexibility applies to verbs as well. In Incorporating Coercive Constructions into a Verb Lexicon, Bonial et al. et al. moved beyond static verb definitions by introducing constructional profiles. These are probabilistic measures that predict how likely a verb is to be coerced into a new meaning by the construction surrounding it. For example, the verb blink does not inherently inherently imply motion, but in the phrase "blinking the snow off," the caused-motion construction coerces the verb into a novel semantic role.

Jaime G. Carbonell argued in Metapher - A Key to Extensible Semantic Analysis and Towards a Self-Extending Parser that metaphor is not an irregular linguistic glitch, but a structured, predictable system. By codifying a core set of generalized mappings, a self-extending parser can relax literal constraints on the fly, mapping physical actions to abstract concepts seamlessly when standard constraints are intentionally violated.

Deconstructing Fixed Ontologies

The brittleness of traditional NLP often stems from its reliance on fixed, expert-heavy taxonomies. Conventional Semantic Role Labeling (SRL), for instance, forces words into abstract, mutually exclusive categories like AGENT or PATIENT.

To capture the inherent fuzziness of human language, He et al. introduced QA-SRL, a framework that replaces abstract linguistic codes with intuitive, natural language question-answer pairs (e.g., asking "Who did what to whom?"). By substituting rigid schemas with fluid natural language descriptions, the framework captures nuances that abstract labels miss.

Taking this a step further, White et al. developed the Universal Decompositional Semantics (UDS) dataset. Rather than forcing a word into a single categorical bucket, UDS decomposes semantic roles into a set of continuous, scalar properties. This allows systems to handle instances that overlap multiple categories or fall outside predefined boundaries entirely.

Categorical SRL (Brittle):
"The wind opened the door." -> Wind = AGENT? (But wind isn't intentional) -> Classification Failure

Decompositional Semantics (Flexible):
Wind -> [Agressor: 0.8, Volition: 0.0, Causation: 1.0] -> Nuanced Understanding

Similarly, Kurohashi et al. bypassed rigid hand-coded taxonomies for Japanese noun phrases by dynamically reading dictionary definitions to uncover hidden semantic roles. By extracting context-dependent relationships on the fly, their system avoids the boundary errors common to fixed classification schemes.

Pragmatic Pointers and Multimodal Flexibility

In real-world conversation, semantic constraints are often bypassed entirely for pragmatic reasons. In Donnellan's Distinction and a Computational Model of Reference, Amichai Kronfeld demonstrated that speakers frequently use descriptions as pointers rather than literal definitions. If a speaker refers to "the man drinking a martini," the reference succeeds even if the glass actually contains water. Rigid semantic matching fails here because the target object's attributes do not perfectly fit the description; a flexible model, however, recognizes the description's referential intent over its literal truth.

This necessity for flexibility is echoed by Kok-Wee Gan in Integrating Word Boundary Identification with Sentence Understanding and by Velardi et al. in their work on the PETRARCA system. Both emphasized that when a system encounters metonymies or idioms that violate standard selectional restrictions, the violation should not trigger a processing failure. Instead, it should act as a diagnostic signal, triggering a concurrent, flexible interpretation mechanism to resolve the non-literal meaning.

This challenge extends directly into modern multimodal tasks like visual grounding. When a model must locate an object in an image based on a natural language phrase, rigid linguistic templates fail if the description is abstract or poetic. As Yang et al. and Lyu et al. demonstrated, breaking language down into fixed syntactic categories struggles with complex, real-world expressions. By implementing mutual guidance mechanisms—where visual features and language embeddings continuously refine each other—modern architectures achieve flexibility, allowing systems to interpret highly irregular or figurative descriptions within concrete visual scenes.

Key Foundational Work

The Generation of Continuous Semantic Constraints from Semantic Propositions (1977) — contributed 2 concept elements to this direction.
Knowledge Structures and Language Boundaries (1977) — contributed 1 concept elements to this direction.
Understanding Without Proofs (1973) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_figurative_language_interpretation_seman_3d066d79,
  title = {figurative language interpretation (semantic constraint flexibility)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/figurative-language-interpretation-semantic-constraint-flexibility/rdr_3d066d792a1c4117a26e6a594adaf3b8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_figurative_language_interpretation_seman_3d066d79_v1,
  title = {figurative language interpretation (semantic constraint flexibility)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/figurative-language-interpretation-semantic-constraint-flexibility/rdr_3d066d792a1c4117a26e6a594adaf3b8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake