Dependency Based Grammar Error Detection For Low Resource Languages

Abstract

Abstract This dissertation introduces a end-to-end, data-driven framework for automated grammar error detection (GED) in Sinhala, a low-resource, free-word-order language. We begin by constructing a 400 sentence Universal Dependencies (UD) treebank via a hybrid LLMand- expert annotation pipeline. Using custom 300 dimensional FastText embeddings, we train a graph-based UUParser with targeted data augmentation and cross-lingual transfer from Hindi. Our final parser achieves an Unlabeled Attachment Score (UAS) of 71.37% and Labeled Attachment Score (LAS) of 55.42%, and attains sentence-level parse accuracy of 82% on a standard correct corpus outperforming a leading CFG-based parser (60%) while retaining 64% accuracy on free-word-order variants . Building on this, we generate a synthetic GED corpus of 10,000 sentences covering five error types. We engineer multi-level token features—pretrained word embeddings, POS embeddings, morphological concatenations, dependency relation embeddings, and syntactic n-grams— and train a BiLSTM classifier. The combined model delivers 80% overall classification accuracy . On a 200 sentence standard evaluation set, it correctly classifies 82% versus 60% for prior CFG-based methods, and it generalizes to free-word-order GED with 64% accuracy. Our contributions include a UD treebank for Sinhala, an optimized dependency parser pipeline, the first dependency-enhanced GED classifier for Sinhala, and a synthetic error corpus. These results confirm that combining hierarchical dependency features with surface-level features significantly boosts GED in challenging low-resource, morphologically rich, free-word-order settings.