FrameNet: Frame Semantic Annotation in Practice

Abstract

Beginning with an overview of the theory of Frame Semantics as developed by Charles Fillmore and colleagues, this article details the annotation of English sentences by the FrameNet Project based on this theory. Fillmore’s lexical semantics theory asserts that the meanings of most words are understood via the semantic frames they evoke; e.g. arrest, apprehend, apprehension, bust, and nab can all evoke the Arrest frame, with its associated frame-specific semantic roles: Suspect, Authorities, Offense, and Charges. Thus, They were busted for shoplifting by three plainclothes policemen would be labeled to show that bust is the frame-evoking expression, they fills the Suspect role, for shoplifting is the Offense, and by three plainclothes policemen represents the Authorities. Combining multiple annotations of this type creates a picture of the valence (valency) patterns of the lexical unit (word sense) and the semantic frame. The resulting database contains more than 200,000 manual annotations of 13,500 lexical units in 1,200 semantic frames. Expanding from the original goal of lexicography, the team has annotated a number of texts “fully”, i.e. labeling all the frame-evoking elements and the phrases that fill their semantic roles, providing a rich representation of the lexical semantics of the entire text. Automatic semantic role labeling systems trained on FrameNet can label a wide range of texts with increasing accuracy for NLP research and applications. The author describes current limitations and possible extensions of this methodology and how the practice of manual annotation informs the development of the theory.