📄 Uniformisation de corpus anglais annotés en sens (Unification of sense annotated English corpora for word sense disambiguation)

Summary

English word sense disambiguation (WSD) has roughly fifteen sense-annotated corpora. They come in different file formats and use different releases of Princeton WordNet, so combining them takes real effort. Habit also splits them unevenly: systems are scored only on corpora built for evaluation, with no scientific reason to exclude the others. The authors gather twelve corpora into one simple XML format. They convert every sense annotation to WordNet 3.0 using mapping tables built with the method of Daudé et al. (2000). The collection includes training-oriented corpora such as SemCor, DSO, OMSTI, MASC, OntoNotes and the WordNet gloss corpus, plus the six SemEval/SensEval evaluation sets. All corpora are treated as equals rather than split into train and test. The contribution is a resource, with no experimental scores. It includes Java code to read and write the format, the conversion scripts, and the corpora in original and unified form where licenses allow. This makes larger-scale WSD evaluation possible, for example by rotating each corpus in turn as the test set.

Key points

• The resource covers 12 corpora, compared with 7 in the similar effort by Raganato et al. (2017). That work also keeps a fixed split into training and evaluation sets.
• Corpora and their original WordNet versions:
  • SemCor: WordNet 1.6
  • DSO: WordNet 1.5
  • WordNet gloss corpus: WordNet 3.0
  • OMSTI, MASC and OntoNotes: WordNet 3.0
  • Six SemEval/SensEval English tasks
• All senses are mapped to Princeton WordNet 3.0 through version-to-version mapping tables based on Daudé et al. (2000).
• The XML format has five nested elements: corpus, document, paragraph, sentence and word.
  • Any element may carry arbitrary attributes.
  • Only surface_form on a word is required.
  • Standard attributes are pos, lemma, and a version-tagged sense key such as wn30_key.
• Code, original data and converted data are on GitHub (getalp/WSD-TALN2017-Corpus-Vialetal). Copyrighted corpora can be rebuilt from their originals using the shipped scripts.

Interesting ideas

• I1. Treating every annotated corpus as both potential training and test material allows a leave-one-corpus-out rotation, which tests generalization across domains instead of on one fixed benchmark.
• I2. A minimal schema with one mandatory field and open optional attributes keeps a format easy to read while still preserving every detail the source corpora provide.
• I3. When licenses block redistribution, shipping the conversion code lets users rebuild the unified data from originals they obtain themselves.
• I4. Mapping old sense labels to a single current inventory through structural mapping tables makes legacy annotations usable together.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Rotate our treebanks and silver sets so each in turn is held out, revealing parser or frame-extractor weaknesses that only appear on unseen sources.
• Encoding the inventory version inside the attribute name (e.g. wn30_key) lets annotations from several lexicon releases sit side by side without ambiguity. — (world): Tag compiled WordNet/FrameNet/Wikidata links with their source version, so ontology upgrades can be migrated and audited rather than silently overwritten.
• I2 (tree): An internal interchange format for converted UD dialects could require only token form and keep dialect-specific features as optional attributes, so no source information is lost.
• I3 (tree): Distribute treebank-conversion scripts instead of converted restricted treebanks, keeping the pipeline reproducible and legally clean.
• I4 (world): Remap older-WordNet sense annotations onto our WordNet 3.0-based ontology, enlarging the evidence behind each concept and verb class.