VerbNet

VerbNet organizes over 6,000 English verb senses into a hierarchical taxonomy of approximately 270 top-level classes, refined into subclasses. This structure captures shared syntactic and semantic behavior among verbs. For example, the class give-13.1 groups verbs like give, hand, and pass together, while throw-17.1 groups throw, toss, and fling. Each class inherits properties from its parent, enabling generalization across semantically related verbs.

Each VerbNet class specifies a set of syntactic frames—surface-level patterns that describe how a verb's arguments are realized. These frames capture diathesis alternations such as:

• Transitive: "Sam broke the window"
• Passive: "The window broke"
• Dative Shift: "Sam gave Mary the book" / "Sam gave the book to Mary"
• Resultative: "Sam hammered the metal flat"

Frames are defined with constituent types (NP, PP, ADJP) and grammatical functions, providing a precise mapping between syntax and semantics.

VerbNet maps each syntactic argument position to a thematic role using a fixed inventory of approximately 20 roles, including:

• Agent: The volitional causer of an event
• Patient: The entity undergoing change
• Theme: The entity in motion or being located
• Recipient: The endpoint of a transfer
• Instrument: The tool used to perform an action

These roles are consistent across frames within a class, enabling semantic role labeling systems to generalize from known verbs to unseen members of the same class.

VerbNet encodes selectional preferences that constrain the ontological types of arguments a predicate can accept. These restrictions are expressed using a set of semantic predicates referencing an ontology of types:

• +animate: Requires a living entity (e.g., Agent of murder-42.1)
• +concrete: Requires a physical object (e.g., Patient of break-45.1)
• +organization: Requires an institutional entity

These constraints prevent semantically anomalous parses and improve argument identification accuracy in SRL systems.

VerbNet represents verb meaning using a subeventual decomposition based on Davidsonian event semantics. Each class specifies a temporal sequence of subevents with predicates like:

• CAUSE: Agentive causation (e.g., break: CAUSE(Agent, BECOME(broken(Patient))))
• MOTION: Directed movement (e.g., throw: CAUSE(Agent, GO(Theme, TO(Location))))
• STATE: Resultant condition (e.g., love: STATE(Experiencer, Stimulus))

This decomposition links directly to FrameNet and PropBank representations, enabling cross-resource mapping.

VerbNet is explicitly mapped to both PropBank and FrameNet, forming a unified lexical resource ecosystem:

• VerbNet-PropBank mapping: Each VerbNet class links to PropBank rolesets via the SemLink project, connecting thematic roles to verb-specific numbered arguments (Arg0, Arg1, etc.)
• VerbNet-FrameNet mapping: VerbNet classes align with FrameNet semantic frames, linking thematic roles to frame elements

This interoperability allows SRL systems trained on PropBank to output VerbNet thematic roles, providing richer semantic representations.