Semantic Frames

Charles Fillmore introduced it in "Frame Semantics" (1982). The frame contains roles: Buyer, Seller, Goods, Money. English verbs profile subsets — "buy" foregrounds Buyer and Goods (Mary bought the book); "sell" foregrounds Seller and Goods (John sold the book); "pay" foregrounds Buyer and Money (Mary paid twenty dollars); "spend" foregrounds Buyer and Money on Goods (Mary spent twenty dollars on the book); "cost" foregrounds Goods and Money (the book cost twenty dollars); "charge" foregrounds Seller, Buyer, Money. All six verbs share one frame; each profiles different participants. Without the frame, the verbs are arbitrary; with it, they are systematic.

Componential analysis (Katz and Fodor, 1963) decomposes word meaning into atomic features — bachelor = +HUMAN, +MALE, +ADULT, -MARRIED. Fillmore argued this misses the frame — "bachelor" presupposes a culture with marriage as a default expectation. Pope John Paul II is technically +HUMAN, +MALE, +ADULT, -MARRIED but not naturally called a bachelor; a man in a permanent same-sex partnership is questionable. The frame matters more than the features. Fillmore's 1975 paper "An Alternative to Checklist Theories of Meaning" laid out this critique. Frame semantics is holistic; componential analysis is decompositional.

FrameNet is a lexical database started in 1997 at the International Computer Science Institute in Berkeley under Fillmore's direction. It catalogs frames (e.g., Revenge, Communication, Self_motion), specifies their core and peripheral frame elements, and documents lexical units (verbs, nouns, adjectives) that evoke each frame, with annotated example sentences. As of 2024 it covers over 1,200 frames and 13,000 lexical units for English. Spanish FrameNet, Japanese FrameNet, German SALSA, and Chinese FrameNet replicate the structure. FrameNet powers semantic role labeling — extracting Who-did-What-to-Whom from text, a core NLP task.

Roger Schank and Robert Abelson's scripts (Scripts, Plans, Goals, and Understanding, 1977) and Marvin Minsky's frames (1974) were AI knowledge structures. Schank's restaurant script details the temporal sequence — entering, ordering, eating, paying, leaving. Fillmore's frames are more abstract — Commercial_Transaction does not specify temporal order; it specifies role structure. Frames may be embedded in scripts. Fillmore explicitly cited Minsky and Goffman (Frame Analysis, 1974) but distinguished his usage. Frame semantics targets lexical meaning; scripts target event sequences. Both contributed to modern semantic representation in NLP.

Adele Goldberg's Constructions: A Construction Grammar Approach to Argument Structure (1995) integrated frames with syntactic constructions. The Ditransitive Construction (X CAUSES Y to RECEIVE Z) maps onto the Transfer frame. The Caused-Motion Construction (X CAUSES Y to MOVE Z) maps onto motion frames. Constructions and frames pair form with meaning at multiple grain sizes. This is the foundation of Cognitive Grammar (Ronald Langacker), Construction Grammar (Goldberg, Croft), and Sign-Based Construction Grammar (Sag, Boas). Frames provide the conceptual anchors; constructions provide the syntactic patterns.

Core elements are conceptually necessary participants — for the Revenge frame, the Avenger, Offender, Punishment, and Injury are core. Peripheral elements add detail — Time, Place, Manner, Reason. The distinction matters for frame identification — sentences must instantiate at least the core elements to evoke the frame. FrameNet uses semantic types (Sentient, Physical_object, etc.) to constrain element fillers. Computational systems like SEMAFOR (Dipanjan Das, 2010) and recent transformer-based semantic role labelers (Strubell et al., 2018) leverage these distinctions to label arguments in text.

Different verbs evoking the same frame foreground different participants and adopt different perspectives. "Risk" evokes a frame with Actor, Valued_object, Action, Harm. "Bertha risked her life rescuing the child" perspectivizes from Bertha's viewpoint; "the child cost Bertha her life" reorders the same frame. Fillmore and Beryl Atkins's 1992 paper on "risk" showed that verbs and nouns offer alternative perspectives onto a shared frame. Translation across languages often requires perspective shifts — the frame is shared, but lexical choices differ. This complicates word-for-word translation and motivates frame-based machine translation.