French Linguists Release GLeMM, a 2.3M-Pair Open Word-Meaning Lexicon
A fully automated pipeline turns Wiktionary into a 2.3-million-pair derivational lexicon spanning seven European languages, with semantic annotations attached.
Takeaways
- GLeMM releases 2.3 million derivational word pairs across seven European languages, built automatically from Wiktionary
- Roughly half the pairs carry semantic descriptions pulled from dictionary definitions, a first at this scale
- Extraction uses FAPinette, a three-stage method based on formal analogies and regex alternation patterns
- Manual sampling shows 94 to 100 percent pair accuracy across English, French, and Italian samples
- Covers atypical phenomena like back-formation, parasynthesis, and overabundance that MorphyNet misses
- Available under CC-BY-4.0 at the CNRS git repository
A team of French computational linguists has released GLeMM, a derivational morphology resource that dwarfs most existing lexicons and, for the first time at this scale, ships with semantic descriptions attached to a large chunk of its entries. The dataset is built automatically from Wiktionary, covers seven European languages, and is available on a public git repository under CC-BY-4.0.
Derivational morphology studies how words like scarce become scarcity, or why English speakers form loneliness instead of lonelity. Answers to these questions typically rest on intuition and observations drawn from limited data, even when a wide range of languages is considered. Hypotheses in the field are hard to replicate and harder to test computationally.
Where existing resources run out of road
Manually curated lexicons like CELEX are trusted as gold standards but small and expensive to build. Automatic ones like MorphyNet, UniMorph, and DeriNet scale better, with precision ranging from 85% for DeriNet.ES (Spanish) to 93% for DerivBase-1.4 (German) and 95% for DeriNet.RU (Russian). Almost none of them describe the meaning side of derivational relations, which is exactly what morphologists need to study phenomena like affix rivalry or back-formation.
How GLeMM is built
The core idea is elegant: if word B appears inside the dictionary definition of word A, and the pair also shows a systematic formal pattern shared by many other pairs, then A and B are probably morphologically related. The authors call these