Building a Community-Grounded Spell Checker for Kabyle Using Hybrid NLP: software and language resources

Kabyle (Taqbaylit) is a Berber language of Algeria and its diaspora, with limited computational resources. This deposit contains a spell checker for Kabyle and the language resources it is built on. The system is deterministic and uses no neural language model: each correction is attributable to an identifiable rule or score. It combines a lexicon with per-source attestation, an edit distance weighted by a hand-specified table of Kabyle letter confusions, morphological and part-of-speech information, a statistical language model and a semantic graph, and it decodes over the sentence rather than word by word. It treats the state of annexation, in which a noun changes form according to its syntactic position and both forms are valid words. The output is a ranked set of candidates rather than a single correction. Two generations of the system are included, evaluated under an identical protocol; the second scores higher than the first on all reported measures. The deposit also reports a negative result: arbitration by neural language models was tested on the cases the pipeline left unresolved and did not improve its scores; that code is not included. The provenance and licensing status of each source are recorded in NOTICE.md. Software under the MIT License, data under CC BY-SA 4.0.