Bangla Grammatical Error Detection

This project aims to develop a grammatical error detection model for the Bengali language, capable of identifying various error types such as single-chunk errors, multi-chunk errors, punctuation errors, and spacing errors.

To address this challenge, we fine-tuned the ByT5-small model, a multilingual model that utilizes byte-level tokenization, making it highly adaptable to Bengali. Additionally, we implemented a custom greedy decoding algorithm to efficiently insert error indicators into the text.

In the post-processing stage, we applied a heuristic-based approach to handle spacing issues and multiple punctuation errors effectively. Our method achieved edit distance of 1.14.