Verb∋Net

What is in a class

VerbNet (Kipper-Schuler, 2005) builds on Levin's (1993) verb classes, in which verbs are grouped by meaning and by the syntactic alternations they allow. For each class and subclass, Verb∋Net gives:

• the members, the French verbs belonging to the class;
• the thematic roles (Agent, Theme, Destination…), optionally with selectional restrictions such as +animate or +concrete;
• the frames. Each frame has an example sentence, a syntactic formula that maps syntactic arguments to thematic roles, and a semantic formula built from predicates over the event and its participants.

For example, class accept-77 contains accepter, admettre, refuser…, with the frame NP.Agent V que Psubj.Theme illustrated by Luc a accepté que son fils devienne écrivain, and the semantics approve(during(E), Agent, Theme).

How it was built

Verb∋Net was built at CEA LIST and Université Paris Diderot as part of the ANR ASFALDA project, whose goals also included a French FrameNet. The method, described in Pradet, Danlos & de Chalendar (2014), keeps VerbNet's class hierarchy and semantics wherever possible. French verbs and frames come from two existing French lexicons:

• Les Verbes Français (LVF, Dubois & Dubois-Charlier, 1997), a semantic lexicon of about 25,000 entries in 14 classes, 54 subclasses and 248 sub-subclasses;
• the Lexique-Grammaire (LG, Gross 1975; Boons et al. 1976), a syntactic lexicon of about 14,000 entries in 67 tables, where each table groups verbs with the same frames.

1. Translating the members. Each VerbNet class was manually linked to the LVF classes and LG tables that match its meaning. These links are kept in the lvf and ladl attributes of each class. The English members were translated with two bilingual dictionaries (SCI-FRAN-EURADIC and the French Wiktionary). A translation was kept only if it belongs to both the LVF classes and the LG tables, or to whichever of the two was non-empty. In the original study, this produced 4,058 verb–class pairs (2,128 distinct verbs) that were precise and coherent.
2. Adapting the frames. Lexicographers then reviewed every class in a collaborative web editor. They split verbs into subclasses, kept the frames valid in French, and adjusted roles, restrictions and prepositions. Some general principles were applied:
   • no sub-frames: a frame that differs from another only by a missing complement is not recorded, because it can be inferred;
   • a single order of complements: Nora a apporté le livre au meeting is recorded, while the variant with the complements swapped is left implicit;
   • no conative, dative or benefactive alternations, since they do not exist in French;
   • the hierarchy is revised where the two languages differ. Examples are verbs of sending and carrying (no equivalent of classes 11.3 and 11.4) and verbs of change of possession (e.g. with becomes en or de depending on the verb).

This revision

This repository contains a revised version of the resource. Guy Lapalme and Marie-Claude L'Homme prepared it while studying the use of Verb∋Net for text generation. The revision:

• adds VERBENET.xml, which includes every class file through XInclude so that XPath queries can run over the whole lexicon;
• adds a RELAX NG schema, VERBENET.rnc, to validate the files;
• fixes typos in some examples and member verbs;
• makes sure each MEMBER element contains exactly one verb;
• corrects some frames and roles so that they match their examples.

Using the data

The reference data is one XML file per class in the verbenet/ directory. The files follow VerbNet's format, with French-specific syntactic elements (see Format). Each release (latest: 1.0.0) is also available in these forms:

• Zip archive on the release page: XML files, schema, and two JSON Lines tables (one row per frame, one row per verb–class pair). Every release is also archived on Zenodo with a DOI (10.5281/zenodo.23018909);
• Python package: pip install verbenet, which contains the XML files and a small loader;
• Hugging Face dataset aymaralima/verbenet: load_dataset("aymaralima/verbenet", "frames");
• the files of the current version, published with this site under data/.

import verbenet

for row in verbenet.frame_rows():       # roles and members already inherited from parent classes
    if "manger" in row["members"]:
        print(row["class_id"], row["primary"], "|", row["examples"][0])

# every verb of the lexicon, through the aggregate file
xmllint --xinclude --xpath '//MEMBER/@name' verbenet/VERBENET.xml

# the classes a given verb belongs to
xmllint --xinclude --xpath '//MEMBER[@name="manger"]/ancestor::*[@ID][1]/@ID' verbenet/VERBENET.xml

from lxml import etree

lexicon = etree.parse("verbenet/VERBENET.xml")
lexicon.xinclude()
for frame in lexicon.iterfind(".//VNCLASS[@ID='put-9.1']//FRAME"):
    print(frame.find("DESCRIPTION").get("primary"), "|", frame.findtext("EXAMPLES/EXAMPLE"))

Citing

If you use Verb∋Net, please cite:

Quentin Pradet, Laurence Danlos and Gaël de Chalendar. 2014. Adapting VerbNet to French using existing resources. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 1122–1126, Reykjavik, Iceland. ELRA. (PDF)

@inproceedings{pradet-etal-2014-adapting,
    title = "Adapting {V}erb{N}et to {F}rench using existing resources",
    author = {Pradet, Quentin and Danlos, Laurence and de Chalendar, Ga{\"e}l},
    booktitle = "Proceedings of the Ninth International Conference on Language Resources and Evaluation ({LREC}'14)",
    month = may,
    year = "2014",
    address = "Reykjavik, Iceland",
    publisher = "European Language Resources Association (ELRA)",
    pages = "1122--1126"
}

To cite the data itself, use the Zenodo DOI 10.5281/zenodo.23018909, which always resolves to the latest version. Each version also has its own DOI, listed on the Zenodo record.

License

Like the original Verb∋Net release, this version is distributed under the Creative Commons Attribution-ShareAlike 4.0 International license. The original work was funded by the French national research agency (ANR) through the ASFALDA project, reference ANR-12-CORD-0023.