Predicates impose selectional restrictions—constraints on the semantic types of their arguments. Verbs like "eat" require animate subjects and consumable objects; "believe" requires sentient subjects. These restrictions are lexical properties stored in a predicate's argument structure and may reduce to semantic well-formedness conditions (arguments must denote individuals with specific properties).
From your work on semantic linking to syntax, you know that verbs impose requirements on their arguments — not just structural requirements (a subject must appear, an object may appear) but semantic ones. The verb "eat" does not just require two noun phrases; it requires that the subject be animate and the object be something consumable. "The rock ate sadness" violates no syntactic rules but is semantically deviant. These semantic constraints are selectional restrictions: requirements that a predicate places on the semantic types of its arguments.
Selectional restrictions are stored as part of a word's lexical entry — they are properties of individual predicates, not derived from general grammar rules. "Elapse" requires a temporal subject (time elapses, but rocks do not). "Blond" requires a head noun denoting something that can bear hair. "Devour" requires an edible object. Each predicate carries a set of semantic filters that its arguments must pass. If an argument fails a filter, the result is selectionally anomalous — it might be interpretable as metaphor or be felicitous in a fictional context, but in literal use it is odd in a way that purely syntactic violations are not.
The theoretical question is whether selectional restrictions are a separate constraint level or reduce to more general semantic well-formedness. On the semantic features approach, nouns carry bundles of features (±animate, ±human, ±concrete), and verbs specify which values their arguments must carry: "eat" requires [+animate] for its subject. On the type-theoretic approach, predicates impose type requirements: "elapse" selects for an argument of type *time interval*, not type *individual*. Both approaches capture the core intuition — arguments must be the right kind of thing — but they make different predictions about borderline cases and how metaphorical extensions are licensed.
Understanding selectional restrictions illuminates how metaphor works. When we say "time devoured the years" or "anxiety gnawed at her confidence," we deliberately violate selectional restrictions to generate non-literal meaning. The violation is productive precisely because the listener recognizes it as a violation: they know "devour" selects for a consumable object, and when the object is abstract, that mismatch prompts them to construct an analogy. Selectional restrictions are the background against which figurative language figures — knowing the rules is what gives a violation meaning.