The Berkeley FrameNet Project

What it is about

The paper describes the Berkeley FrameNet project, a corpus-based lexicon for English. Its goal is frame-semantic descriptions of several thousand English lexical items, backed by annotated corpus sentences. Each frame gives frame elements, the roles of an event, and each target word has a valence description that links its frame elements to grammatical functions and phrase types. The authors plan at least 5,000 lexical entries with a parallel annotated corpus, for use with WordNet and COMLEX. The project works in stages: a Vanguard prepares frame descriptions, annotators tag example sentences by hand, and a Rearguard writes the dictionary-style entries. The paper gives the workflow and the tools, and no accuracy figures.

Ideas

I1. world/src/frames.rs: map each English verb's arguments in EWT test to frame elements through the dependency labels (subject, object, oblique), and count how often the mapping is consistent with a frame's linking. Measure the match rate per frame on held-out sentences.
I2. Pattern coverage for the school lessons: for each verb in the school silver corpus, choose example sentences that cover each argument pattern (the paper's annotated example set), and report the share of patterns covered per verb in world lessons.
I3. Frame-element inventory: check the frame elements that onto (FrameNet 1.7) lists against the roles the world reader fills, and count the roles with no reader rule. The count is the measure, before any new rule is written.

For Mova

The code is world/src/frames.rs and the FrameNet data in onto. The chapters are languages/en/28-voice-and-passive.md for the linking of arguments, and 21-clause-structure.md for the argument slots. The paper is an early project description, and the FrameNet data it describes has its own licence, so use it for design ideas and tests, not as a source of build data. Measure first: the frame-linking match rate on held-out EWT sentences, two identical runs.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written by Mova (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, materials before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/P98-1013.pdf

Summary

The paper describes Berkeley FrameNet, a corpus-based computational lexicography project funded by the NSF for three years. Its goal is to describe several thousand English nouns, verbs and adjectives in terms of frame semantics. Each frame defines its own set of roles, called frame elements (FE). Unlike thematic roles, these roles are not universal; they are tied to a specific conceptual situation. Subframes inherit the roles of their parent frame; DRIVING, for example, extends TRANSPORTATION. Lexicographers manually annotate sentences from the British National Corpus. Suitable examples are selected by syntactic patterns using a cascaded filter based on CQP. The work is divided into four stages: preparation, subcorpus extraction, annotation and entry writing. A toolkit based on PERL/CGI and SGML was assembled for it. The intended result is a linked database with three parts: a lexicon with valence formulas, a frame database and annotated examples. It is planned to be integrated with WordNet and COMLEX. For NLP, such a resource indicates where in a sentence to look for a word's arguments and which semantic role they play. It later became the basis for automatic semantic role labeling.

Key points

At the time of publication, almost 10,000 sentences had been annotated for about 200 lemmas. They contain more than 20,000 annotated FE tokens; about a dozen frames with 47 named FEs have been described.
The ultimate goal is at least 5,000 lexical units. If current proportions hold, the database will contain about 250,000 annotated sentences and more than half a million FE tokens.
The corpus is the BNC in the POS-tagged and lemmatized version from IMS Stuttgart. Tools: CQP and XKWIC (IMS), Alembic (Mitre) for SGML markup, and the in-house utilities arf and whittle.
Participants have functional roles: the Vanguard describes frames, the Annotators annotate sentences, and the Rearguard writes dictionary entries.
The data are stored in SGML with a separate DTD for each component; the project later plans to move to XML.

Topics: Semantic role labeling

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/980845.980860

@inproceedings{baker1998berkeley,
  title = {{The Berkeley FrameNet Project}},
  author = {Baker, Collin F. and Fillmore, Charles J. and Lowe, John B.},
  booktitle = {ACL-COLING 1998 (Montréal)},
  year = {1998},
  doi = {10.3115/980845.980860},
  url = {https://doi.org/10.3115/980845.980860}
}

References

6 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Charles J. Fillmore and B. T. S. Atkins. forthcoming. FrameNet and lexicographic relevance. In Proceedings of the First International Conference On Language Resources And Evaluation, Granada, Spain, P8-30 May 1998. find
Susanne Gahl. forthcoming. Automatic extraction of subcorpora based on subcategorization frames from a part of speech tagged corpus. In Proceedings o/ the 1998 COLING-A CL conference. find
Institut f'dr maschinelle Sprachverarbeitung IMS. 1997. IMS corpus toolbox web page at stuttgart, http://www.ims.uni-stuttgart.de/~oli/CorpusToolbox/. ↗
John B. Lowe, Collin F. Baker, and Charles J. Fillmore. 1997. A frame-semantic approach to semantic annotation. In Tagging Text with Lexical Semantics: Why, What, and How? Proceedings of the Workshop, pages 18-24. find
Special Interest Group on the Lexicon, Association for Computational Linguistics, April. find
Mitre. 1998. Alembic Workbench web page at Mitre corp. http: //www.mitre.org/resources/centers/advanced\_info/g04h/workbench.html. find