Is a restriction always a hard rule?

Formal systems can impose one, but language applications often need graded preferences and contextual exceptions.

Does a compatible argument prove an event occurred?

No. Compatibility, textual assertion, and external verification are different questions.

Can preferences help resolve ambiguity?

Yes, as one source of evidence alongside syntax, discourse, and the task context.

AI Glossary / Natural language and speech / Advanced

Selectional restriction

Definition

A selectional restriction is a semantic constraint that a particular predicate sense places on an argument, such as the kind of thing that can fill its Theme or Agent role. It concerns compatibility of meanings, beyond grammatical form. Natural language often requires softer preferences, alternative senses, or contextual interpretation rather than unconditional rejection.

What it means

A sentence can have an ordinary grammatical structure while describing an unexpected combination of participants. Selectional restrictions help an interpreter notice that mismatch and consider what the speaker intended.
For an invented miniature language model, suppose one literal sense of drink requires a liquid as its Theme. Water fits that requirement; a granite block does not. This says nothing about whether the water is safe, whether anyone actually drank it, or whether the sentence is true. It also does not cover figurative expressions such as drinking in a view. The restriction belongs to a sense and role, not every appearance of the spelling drink.

How it works

A symbolic implementation identifies a predicate sense, assigns argument roles, and looks up semantic types in an ontology. It then checks whether an argument's type satisfies the specified role constraint, allowing inheritance from narrower types to broader ones.
A statistical implementation can learn which argument classes occur disproportionately with a predicate. These graded selectional preferences can support interpretation without treating every unusual combination as impossible. Counts depend on the corpus and on how words, senses, and classes were identified.
When a mismatch occurs, useful next steps include checking the parse, word sense, reference, or figurative interpretation. Simply deleting the sentence confuses a limitation of the model with a defect in the language.

Three practical examples

A document parser uses participant types to distinguish a physical repair description from a discussion of repairing a relationship. The interpretation must still fit the surrounding text.
A voice interface compares candidate transcriptions partly by whether their predicate and argument meanings fit the task. Semantic plausibility alone cannot prove which sounds were spoken.
A language-learning application explains why a literal reading sounds unusual, while accepting a clearly signaled metaphor instead of treating creative language as a typing error.

Think of it this way

Think of a labeled socket and a collection of plugs. A type check asks whether the selected plug fits that socket. Language adds another possibility: the speaker may be using a different socket, or using the whole arrangement metaphorically. A failed fit starts an interpretation question; it does not finish one.

Worked example

Construct a toy type system with Water and Milk below Liquid, and Granite below Solid. For the chosen literal sense of drink, require Theme to be a Liquid. Evaluate three argument labels: water, milk, and granite. The first two satisfy the stated restriction; the third fails. Thus 2/3 candidates pass this toy check. This is a compatibility rate, not an accuracy score against human judgments.
Now imagine a separate, invented corpus of ten resolved uses of a predicate. Its argument classes occur eight, one, and one times. Their observed conditional frequencies are 0.8, 0.1, and 0.1. If corresponding background frequencies are 0.25, 0.25, and 0.50, the conditional-to-background ratios are 3.2, 0.4, and 0.2.
The first class is enriched relative to its background frequency. These ratios illustrate a preference signal; they are not probabilities and are not Resnik's normalized selectional-association measure. None makes an individual sentence true. A rare but legitimate use remains possible, and the tiny invented sample supports no general language claim.

Illustrative case study

Fictional company ClearPhrase adds semantic checks to maintenance notes. Its first prototype assumes that every object of repair must be equipment. It consequently flags a supervisor's note about repairing trust after a missed deadline.
The team keeps the equipment rule for a narrowly defined literal interpretation but adds an alternative interpretation path and records uncertainty. Reviewers separate incorrect role assignments, unfamiliar equipment names, and legitimate figurative uses. Evaluation includes all three groups rather than only standard machine descriptions.
This fictional case illustrates how a useful expectation becomes harmful when presented as a universal prohibition. It provides no measured claim about product performance.

Common misconceptions and mistakes

A selectional restriction is not a grammatical agreement rule. A noun can agree syntactically while being semantically unexpected. Conversely, semantic plausibility does not repair a malformed parse.
Do not equate typical with permitted, or unusual with false. Negation, quotation, fantasy, metaphor, and specialized terminology can all challenge an overly simple interpretation. An incomplete ontology may also assign the wrong type.
Avoid applying a rule to a word without specifying its sense and argument role. In particular, subject position and Agent role are not interchangeable across all constructions. Treat learned preferences as evidence conditioned on training data, not universal standards of acceptable expression.

Business use

Selectional information can help prioritize ambiguous extractions for review and explain certain interpretation choices. A practical evaluation should measure missed valid expressions as well as detected mismatches. Keep the original sentence, proposed roles, chosen sense, and reason for the flag available to reviewers. This makes it possible to distinguish an actual extraction problem from a restrictive vocabulary model.

Check your understanding

Why does granite fail the toy drink check?

Answer: Its assigned type is Solid rather than a subtype of Liquid.

Why is the ratio 3.2 not a probability?

Answer: It divides a conditional frequency by a background frequency and can exceed one.

What should happen before rejecting a surprising sentence?

Answer: Reconsider its sense, roles, reference, context, and possible figurative reading.

Related terms

Learn next

Create a small sense-specific role inventory, then test literal, figurative, negated, and ambiguous examples. Compare a hard type check with a graded preference score, documenting which errors each approach introduces.

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 22, section 22.7, selectional restrictions and preferences: https://web.stanford.edu/~jurafsky/slp3/22.pdf Resnik, Selectional Preference and Sense Disambiguation, 1997, corpus-based preference modeling: https://aclanthology.org/W97-0209.pdf The type system, counts, ratios, and company scenario above are original illustrations. Review note: Chapter 22 could not be retrieved; Resnik’s primary discussion was consulted.

Previous: Selection problem Next: Selective Adherence Bias