📄 The FrameNet Database and Software Tools

Summary

FrameNet tackles the absence of an English lexicon that ties word senses to the ways their semantic participants show up in syntax. Based on Frame Semantics, it treats each lexical unit (a lemma paired with one sense) as evoking a frame, a conventionalized scene whose participants are frame elements (FEs). It then records how those FEs surface as phrase types and grammatical functions. Instead of annotating running text, lexicographers pick corpus sentences from the BNC and LDC news data that cover the full range of a word's combinatory patterns. They also mark core FEs that are left unexpressed. This demo paper presents the resulting database and the in-house toolchain: a Java frame editor and annotation GUI built on MySQL, web reports, the FrameSQL search interface, and a rule-based FE pre-labeler. The pre-labeler is to be compared with Gildea's learned role labeler. The data, roughly 300 frames and 5,000 lexical units, are distributed as XML. The resource matters because it gives lexicographers and NLP systems corpus-attested semantic-role data, for example for automatic semantic role labeling.

Key points

• Database scale: about 300 semantic frames and roughly 5,000 lexical units, with over 20 annotated corpus examples per lexical unit on average. Frames are linked by inheritance (ISA) and composition (PART-OF).
• Corpora: BNC and LDC English newswire, with plans to add the American National Corpus. Corpus Workbench (IMS Stuttgart) handles search and subcorpus creation. Phase-one annotation used MITRE's Alembic Workbench, later replaced by a custom Java GUI over MySQL.
• Omitted core FEs fall into three types: constructional (e.g., imperative subjects), existential (e.g., unstated objects of eat, bake) and anaphoric (recoverable from discourse).
• Multi-layer annotation supports overlapping FE labels, mismatches between syntactic and semantic constituents, and several targets per sentence. A new XML format reflects this, and RDF via DAML+OIL is being added.
• FrameSQL (Hiroaki Sato, Senshu University) supports searches by frame, lemma, FE, phrase type, grammatical function and head noun. A cascaded rule filter (e.g., valence patterns for tell) pre-labels FEs so annotators mostly approve or correct its output.