Assessing learning materials: hybrid vs Large Language Model-based generation of grammar exercises
Résumé
In this paper, we compare two methods for automatic generation of grammar exercises from authentic texts: (i) a hybrid method where texts are first fed to a neural syntactic parser in order to get morpho-syntactic annotations, from which grammar concepts (e.g. direct objects of verbs) are automatically extracted using handcrafted rules (and then turned into grammar questions), and (ii) a Large Language Model (LLM) -based method where grammar concepts are extracted directly from texts using few shot learning. Concept extraction is evaluated against manually annotated data for French coming from the SEQUOIA treebank, using grammar concepts from French elementary and middle school curricula. Preliminary results show that hybrid approaches offer a viable alternative to pre-trained LLMs in the context of grammar exercise generation.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |