Selectional Preferences

Selectional preferences are formalized as type constraints on argument positions. A predicate like 'drink' selects for arguments of type liquid or beverage, while 'eat' selects for solid food. These constraints operate at the ontology level, not the lexical level—'drink' accepts 'water', 'coffee', or 'smoothie' because they share the relevant semantic type, not because they appear in a fixed list.

• Violation detection: 'The rock drank the water' violates the agent selectional preference for 'drink' (requires animate agent)
• Metaphor resolution: 'The engine drank gasoline' extends preferences metaphorically
• Type coercion: Some constructions force reinterpretation of argument types to satisfy preferences

Modern systems learn selectional preferences automatically from large text corpora using distributional methods. Rather than hand-coding type hierarchies, models infer that 'drink' prefers liquid-type arguments by observing co-occurrence patterns across millions of sentences.

• Vector-space models: Compute the semantic similarity between a predicate's typical arguments and candidate arguments
• Neural language models: BERT and similar architectures implicitly encode selectional preferences in their attention patterns
• Resnik's selectional association: A classic information-theoretic measure quantifying the strength of a predicate's preference for a semantic class

Selectional preferences apply differentially across thematic roles. A single predicate imposes distinct constraints on its agent, patient, instrument, and other arguments.

• Agent role: Typically requires animacy and volition ('John broke the window' vs. 'The storm broke the window')
• Patient role: Requires affectedness and often specific physical properties ('break' selects for fragile objects)
• Instrument role: Requires tool-like properties and functional compatibility ('cut' accepts 'knife', 'scissors', but not 'hammer')
• Location role: Requires spatial containment or surface properties ('put' selects for container or surface locations)

When selectional preferences are violated, language processing systems employ repair strategies to maintain coherence. These mechanisms are critical for handling creative language, metaphor, and domain adaptation.

• Type coercion: The argument is reinterpreted to satisfy the predicate's requirements ('begin the book' coerces 'book' to an event-reading like 'reading the book')
• Metonymy resolution: 'The ham sandwich wants his check' resolves 'ham sandwich' to the customer who ordered it
• Accommodation: In novel contexts, the preference itself may be temporarily relaxed or extended ('The AI reasoned about the problem' extends 'reason' to non-human agents)

Selectional preferences serve as a soft constraint in SRL systems, improving argument identification and classification accuracy. When a parser encounters 'The hammer broke the window', selectional preferences help disambiguate whether 'hammer' is an instrument or agent.

• Argument pruning: Candidates that violate strong selectional preferences are filtered before classification
• Role disambiguation: Preferences help distinguish between semantically similar roles (e.g., Agent vs. Instrument)
• Implicit argument recovery: Strong preferences enable inference of unexpressed arguments ('John ate' implies food as patient)
• Cross-lingual transfer: Selectional preferences often generalize across languages, aiding low-resource SRL

Lexical resources explicitly encode selectional preferences through semantic class restrictions on thematic roles. VerbNet organizes verbs into hierarchical classes, each specifying the ontological types required for each argument position.

• VerbNet classes: The 'break-45.1' class specifies patient must be a physical object with fragility properties
• FrameNet frame elements: The 'Ingestion' frame requires an Ingestibles element of semantic type comestible
• Intersective constraints: Multiple simultaneous preferences (e.g., 'devour' requires both edible patient and animate agent) are combined through type intersection

The generalized semantic categories that define a noun phrase's relationship to the main verb. Selectional preferences constrain which entities can fill these roles.

• Agent: The volitional initiator of an action
• Patient: The entity undergoing the action's effect
• Theme: The entity that is moved or located
• Instrument: The tool used to perform an action

A verb like devour selects an animate Agent and an edible Patient, encoding its preferences directly into its thematic role expectations.

The linguistic framework representing a sentence as a predicate (typically a verb) and its associated arguments. Selectional preferences operate at this structural level, governing the semantic compatibility between the predicate and each argument slot.

• Defines 'who did what to whom, when, where, and how'
• Selectional restrictions are a key constraint on argument realization
• Violations produce semantic anomaly (e.g., The rock ate the idea)

A hierarchical verb lexicon that explicitly encodes selectional preferences alongside syntactic frames and thematic roles. Verbs are organized into classes sharing semantic and syntactic behavior.

• Maps syntactic frames to semantic representations
• Specifies ontological type restrictions on arguments
• Example: The eat-39.1 class restricts the Agent to [+animate] and the Patient to [+comestible]
• Enables generalization across semantically related verbs

A lexical database organized around semantic frames—conceptual structures describing situations, objects, or events. Each frame defines frame elements (roles) that carry implicit selectional constraints.

• The Ingestion frame expects an Ingestor (animate) and Ingestibles (food/drink)
• Selectional preferences emerge from frame semantics rather than individual lexical entries
• Supports cross-lingual preference modeling through frame alignment

The SRL sub-task of assigning a specific semantic role label to a previously identified argument. Selectional preferences serve as critical features for this classification step.

• Neural models learn preferences from co-occurrence statistics
• BERT-based SRL captures contextualized selectional constraints
• Helps disambiguate roles when syntax alone is insufficient
• Example: In The chef cooked the meat, the preference for an animate Agent helps classify chef correctly

The task of converting natural language into formal meaning representations. Selectional preferences guide parsers toward semantically coherent interpretations by filtering implausible argument assignments.

• AMR Parsing uses preferences to resolve ambiguous role assignments
• Preferences reduce the combinatorial search space of possible parses
• Violations signal metaphorical or non-literal language use
• Critical for downstream reasoning systems that require coherent predicate-argument structures