Mei drank the water is normal. Mei drank the sandwich is strange. Mei drank the idea is stranger still.
The syntax can remain legal: subject + verb + object. The literal meaning fails because drink normally expects a drinkable object.
This meaning-based compatibility is described through selectional restrictions or selectional preferences.
Words do not combine only because grammar permits the structure; their meanings also place expectations on the kinds of entities that can fill surrounding roles.
Syntax can be correct while meaning is wrong
The stone ate the sandwich. The sentence has a grammatical skeleton, but a stone cannot normally fill the eater role.
Verbs create participant expectations
Eat normally expects an animate eater and edible object. Read expects an interpreter and text/sign. Repair expects a capable agent and something restorable.
Selectional restrictions differ from semantic frames
A semantic frame tells us the roles in an event. A selectional restriction tells us what kinds of entities can plausibly fill those roles.
They also differ from collocation
Strong tea versus powerful tea is mainly collocational preference. Drink tea versus drink a chair is semantic compatibility.
Restrictions are preferences, not absolute laws
The child ate paper is unusual but physically possible. Real language permits rare events, fiction, technical extensions and metaphor.
Metaphor deliberately violates defaults
The flood swallowed the road. A flood has no mouth. The selectional mismatch tells the reader to search for a figurative interpretation: the road disappeared beneath the water.
Personification works similarly
The city never sleeps. Cities are not biological sleepers. The predicate invites a personified reading: continuous activity.
Semantic change can relax old restrictions
Grasp the handle extends to grasp the concept. Once a metaphorical extension becomes conventional, the former mismatch feels ordinary.
World knowledge matters
The microscope diagnosed the patient may blur agency. A microscope supplies evidence; a clinician or authorised diagnostic system normally performs the diagnosis. Semantic compatibility includes knowledge of real-world capability.
Scientific verbs have strong expectations
The thermometer catalysed the reaction is semantically wrong unless a genuine catalytic mechanism is explained. Subject knowledge and language precision meet in the same sentence.
Selectional restrictions expose bad synonym replacement
The medicine reduced the pain is natural. The medicine shrank the pain is grammatically shaped but semantically odd because shrink strongly prefers size-like change rather than pain intensity.
Argument structure and semantic selection are different
Argument structure tells us how many grammatical roles a verb takes. Selectional restriction tells us what semantic types can fill those roles.
AI can generate selectional failures
The thermometer prescribed antibiotics is fluent English syntax with an impossible institutional capability. This kind of error can be detected by asking whether the participant type can perform the predicate.
A practical compatibility test
- Identify the predicate.
- List its semantic roles.
- Ask what kinds of entities normally fill them.
- Check whether the actual participants fit.
- If not, test for metaphor, metonymy or technical extension.
Singapore relevance
Students learn Science, Mathematics, Humanities and technology through English. A sentence can be grammatically polished and still contain the wrong actor, object or mechanism. Selectional reasoning keeps language connected to reality.
Teacher-friendly diagnosis
Grammar-only checking: syntax accepted as sufficient.
Collocation/selection confusion: conventional preference confused with semantic incompatibility.
Literalism: metaphorical mismatch rejected automatically.
Thesaurus failure: synonym inserted despite incompatible event structure.
World-knowledge failure: participant assigned a capability it does not have.
Hilary Mantel as a high-level literary lens
A strong sentence knows who can act, what can be affected and which object belongs to which process. The transferable lesson is coherent agency unless figurative language deliberately bends it.
Connections eduKateAI can learn
predicate → semantic roles → expected participant types
syntax legality ≠ semantic compatibility
selectional mismatch → error OR figurative reinterpretation
semantic extension → old restrictions can broaden
AI fluent sentence → capability/participant audit
The public principle is that predicates carry meaning-based expectations about the entities around them, so grammatical well-formedness alone does not guarantee a coherent sentence.