📄 Un corpus en arabe annoté manuellement avec des sens WordNet (Arabic Manually Sense Annotated Corpus with WordNet Senses)

Summary

OntoNotes Release 5.0 contains the only freely available Arabic corpus with manual word-sense annotation. The authors argue that it is rarely used because its Arabic sense inventory was never connected to Princeton WordNet, unlike the English and Chinese parts. Without that link, the Arabic portion cannot draw on the tools and resources built around WordNet. To fix this, they extended the Arabic sense-annotated data, which consists only of news text, with alignments from OntoNotes senses to Princeton WordNet 3.0. They used a semi-automatic workflow: WordNet senses were extracted for each lemma, linked to the OntoNotes inventory, and then checked by people. The checking took about four person-months. The result covers more than 12,500 annotated verb and noun occurrences, now reachable through WordNet synsets. The extended resource is intended to be freely released. The authors hope it will become a common benchmark for Arabic word sense disambiguation and make it easier to compare and build WSD systems for the language.

Key points

• The base resource is OntoNotes Release 5.0 (Weischedel et al., 2013, LDC2013T19). It was annotated with about 90% inter-annotator agreement for syntax, predicate-argument structure, word sense and coreference.
• Before this work, the Arabic part (news genre only) had sense annotations, but none of them were linked to Princeton WordNet.
• The extended corpus has 12,524 annotated occurrences: 3,990 verbs and 8,534 nouns.
• These occurrences cover 261 unique lemmas (150 verbs, 111 nouns) and 1,105 unique senses (642 verbal, 463 nominal).
• The authors added 5,558 unique WordNet 3.0 correspondences (4,182 for verbs, 1,376 for nouns), at a cost of about four person-months of semi-automatic work.

Interesting ideas

• I1. Mapping is done once per sense inventory rather than once per token, so 261 lemmas' worth of links automatically relabel more than 12,500 corpus occurrences, which spreads the manual cost very efficiently.
• I2. Linking an isolated local resource to a hub resource such as Princeton WordNet is presented as the main way to make it useful, since the hub's ecosystem then becomes available.
• I3. The links are many-to-many: about 1,105 senses map to 5,558 synsets, roughly five per sense. This means one inventory sense often corresponds to a group of finer WordNet synsets rather than a single one.
• I4. The workflow combines automatic candidate extraction with human verification and reports the human cost explicitly (four person-months).
• I5. A corpus that already carries parallel layers (syntax, predicate-argument structure, coreference, senses) becomes much more valuable once one of those layers is anchored to a shared ontology.

For Mova

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

• I1 (world): when tying our own verb classes or frame senses to WordNet, we should align the inventory entries and let every annotated token inherit the link, instead of tagging text by hand.
• I2 (world): every compiled knowledge table (measures, verb classes, ontology nodes) should carry explicit WordNet/Wikidata IDs, so that new external resources plug in without remapping.
• I3 (frames): treat a frame or sense label as a set of WordNet synsets rather than a single ID, so that role and sense decisions stay explainable while tolerating differences in granularity.
• I4 (learn): when we build silver sense or frame links, we can generate candidates automatically, have an LLM teacher or a human confirm them, and log the verification effort to judge whether the method scales.
• I5 (answer): an evaluation set that links UD trees, frames and WordNet senses on the same sentences would let us test the whole text → answer chain against a single gold standard.