FrameNet

A semantic frame is a script-like conceptual structure describing a particular type of situation, object, or event along with its participants. For example, the Revenge frame evokes an Avenger, an Offender, an Injured Party, and a Punishment. Words like avenge, retaliate, and vengeance all evoke this same frame, unifying synonyms and related terms under a single cognitive representation rather than treating them as isolated dictionary entries.

Frame Elements (FEs) are the semantic roles specific to each frame, analogous to but more granular than generalized thematic roles. In the Commerce_buy frame, core FEs include:

• Buyer: The participant exchanging money
• Seller: The counterparty providing goods
• Goods: The item being transferred
• Money: The payment medium Unlike abstract theta-roles, FEs are defined relative to the specific scenario, providing richer semantic constraints for natural language understanding.

A Lexical Unit (LU) is a pairing of a word with a specific sense that evokes a particular frame. The verb break maps to multiple LUs: Break_off evokes the Removing frame, while Break_down evokes the Cause_to_make_noise frame. FrameNet documents over 13,000 lexical units across more than 1,200 frames, with each LU annotated with example sentences showing its syntactic realizations and the frame elements it governs.

FrameNet records valence patterns—the combinatorial properties specifying how frame elements are realized syntactically for each lexical unit. For the LU give.v in the Giving frame, valence patterns include:

• Donor as External NP Subject
• Recipient as NP Object or PP[to]
• Theme as NP Object or PP[of] These patterns bridge semantics and syntax, enabling parsers to map surface grammatical relations to underlying frame-semantic structures.

Frames are organized in a relational network through eight types of frame-to-frame relations:

• Inheritance: A child frame elaborates a parent (e.g., Revenge inherits from Rewards_and_punishments)
• Subframe: A frame is a component of a complex event (e.g., Arrest is a subframe of Criminal_process)
• Precedes: Temporal ordering between subframes
• Perspectivization: Different viewpoints on the same scene
• Using, Causative_of, Inchoative_of, and See_also This ontology enables inference and generalization across related scenarios.

Beyond the lexical database, FrameNet includes a full-text annotation corpus of continuous documents annotated with frame-semantic layers. Annotators label each sentence's targets (frame-evoking words), their frame elements, and the spans of text filling those roles. This corpus provides training data for automatic semantic role labeling systems and demonstrates how frames interact across discourse, including null instantiation where frame elements are contextually understood but not overtly expressed.