Sorani Gec

Sorani GEC is a grammatical error correction model specifically designed for Central Kurdish (Sorani). The repository contains trained model checkpoints developed as part of an MSc thesis titled 'Agreement-Aware Grammatical Error Correction for Central Kurdish (Sorani): A Morphology-Driven Neural Approach' by Tishko Salah Hawez at the University of Kurdistan Hewlêr. The model addresses grammatical errors in Sorani text, focusing on agreement-aware correction methods. Two model variants are provided, both fine-tuned from the google/byt5-small architecture using a synthetic Central Kurdish grammatical error correction corpus of approximately 27,000 training pairs. The baseline variant operates as a byte-level sequence-to-sequence model without incorporating linguistic features. The second variant, termed 'morphaware,' utilizes the same backbone but augments it with nine morphological features per word, a 33-edge agreement graph, and an auxiliary agreement-prediction loss to enhance its correction capabilities in the context of Sorani's morphological structure. The repository includes model checkpoints, code, and evaluation results, with further resources available via a linked GitHub repository. The tool is released under the MIT license. Sorani GEC is relevant for tasks involving grammatical error correction in Central Kurdish, particularly in low-resource natural language processing contexts. It is positioned as a resource for researchers and practitioners working with Sorani language data or interested in morphology-driven neural approaches to grammatical error correction.