FrameNet, developed at the International Computer Science Institute in Berkeley under the direction of Charles Fillmore, is a lexical resource based on the theory of frame semantics. A semantic frame is a schematic representation of a type of situation, event, or relation, together with its participants and props. For example, the Commercial_transaction frame includes roles for Buyer, Seller, Goods, and Money, and is evoked by words like "buy," "sell," "purchase," "price," and "cost." FrameNet catalogs over 1,200 frames with more than 13,000 lexical units and 200,000 annotated sentences.
Frame Semantics Theory
Core roles: Buyer, Seller, Goods, Money
Non-core roles: Purpose, Place, Time, Manner
Lexical units evoking this frame:
buy.v, sell.v, purchase.n, purchase.v, cost.v, price.n, ...
Frame relations:
Inherits_from: Transfer
Subframe_of: Commerce_scenario
Perspective_on: Commercial_transaction (Buyer vs. Seller)
Fillmore's frame semantics holds that words are understood relative to background knowledge structures (frames) that provide the context for interpretation. Understanding the word "buy" requires activating the Commercial_transaction frame, which specifies that there is a buyer who gives money to a seller in exchange for goods. Different words may evoke the same frame from different perspectives: "buy" foregrounds the Buyer, while "sell" foregrounds the Seller. This perspectival nature of frames distinguishes FrameNet from purely syntactic approaches to argument structure.
FrameNet Annotations
Each FrameNet entry includes annotated example sentences showing how frame elements (roles) are realized syntactically. For instance, in "Kim bought a book from Pat for five dollars," "Kim" fills the Buyer role, "a book" the Goods role, "Pat" the Seller role, and "five dollars" the Money role. These annotations provide training data for automatic semantic role labeling systems. FrameNet distinguishes core frame elements (which are conceptually necessary) from non-core elements (which provide additional information like time, place, and manner).
FrameNet and PropBank are complementary resources for semantic role labeling. PropBank assigns numbered argument labels (Arg0, Arg1, ...) on a verb-by-verb basis, providing broad coverage with shallow semantic distinctions. FrameNet uses semantically meaningful role labels (Buyer, Seller, Goods) organized into frames, providing deeper semantic information but with more limited coverage. Systems trained on PropBank achieve higher coverage, while FrameNet-based systems provide richer semantic representations. Recent work has explored automatic mapping between the two resources.
Computational Applications
FrameNet has been applied to a range of NLP tasks. Frame-semantic parsing systems (Das et al., 2014) automatically identify frames evoked by predicates and label their arguments, extending dependency parsing with semantic role information. In information extraction, frames provide templates for extracting structured events from text. In machine translation, frame-semantic representations can capture meaning that survives translation even when syntactic structure changes. Question answering systems use frame knowledge to match questions to answers that describe the same situation from different perspectives.
FrameNet resources have been extended to other languages, including German, Japanese, Spanish, Chinese, and Brazilian Portuguese, enabling cross-lingual frame-semantic analysis. The Berkeley FrameNet project continues to expand coverage and refine frame definitions. Recent work has explored using large language models to perform frame-semantic parsing, leveraging the models' implicit knowledge of situational semantics to improve coverage beyond what is annotated in FrameNet.