FrameNet

FrameNet is a hand-built, corpus-based lexical resource for English that describes word meaning through the situations words evoke. It is developed at the International Computer Science Institute (ICSI) in Berkeley, California, drawing on the frame semantics of the linguist Charles J. Fillmore. Its starting point is simple but powerful: many words cannot be understood in isolation, only against a frame, a schematic scene with participants. You cannot grasp «to sell» without a seller, a buyer, some goods and a price.

Founded in the late 1990s and now coordinated by Collin Baker, FrameNet turns that intuition into a searchable database, built so that machines —and linguists— can see the situation behind each word.

Frames, frame elements and lexical units

FrameNet rests on three building blocks. Frames are the schematic situations: «Commerce_buy» (a purchase), «Apply_heat» (cooking with heat) or «Being_born». Each frame defines its frame elements, the roles involved: in «Commerce_buy» these are the buyer, the seller, the goods and the money. And each frame gathers the lexical units, the words that evoke it: «buy» and «purchase» trigger «Commerce_buy»; «fry», «bake» and «boil» trigger «Apply_heat». A single scene thus groups verbs, nouns and adjectives that share a background, and every sense is illustrated with real annotated sentences. Release 1.7 of the resource holds more than 1,200 frames, over 13,000 lexical units and close to 200,000 example sentences.

What it is for: semantic roles

FrameNet's classic use is semantic role labeling: given a sentence, working out who did what to whom. In «Ana bought a car from Luis for 9,000 euros», a system trained on FrameNet recognises the buying frame and assigns Ana the role of buyer, Luis that of seller, the car that of goods and the figure that of price. This layer of meaning, deeper than plain syntactic parsing, feeds language understanding, information extraction and question-answering systems.

Against WordNet and in the age of LLMs

FrameNet should not be confused with WordNet. WordNet organises words into synonym sets and hierarchies —what is a synonym of what, what is a kind of what—; FrameNet organises situations and the roles they contain. One maps the vocabulary; the other maps the world the vocabulary refers to. As a hand-curated resource, FrameNet is precise and transparent, but its coverage is limited and its construction slow. Today large language models (LLMs) solve many understanding tasks end to end, without explicit frames, and often outperform it. Even so, frame semantics remains a valuable lens for analysing and explaining how those models —and people— assign the roles that make up meaning.