IndicBERT-CRF Based Chunking Framework for Low-Resource Gujarati Language

Abstract

Chunking is an important NLP task that identifies non-overlapping syntactic phrases such as noun and verb phrases without constructing a full parse tree. For low-resource and morphologically rich languages like Gujarati, chunking remains challenging due to flexible word order, limited annotated data, and insufficient language-specific resources. This study proposes an IndicBERT-CRF framework that combines contextual representations from IndicBERT with CRF-based BIO sequence decoding. Evaluated on a Gujarati corpus of approximately 6,000 sentences and 125,000 tokens, the model achieved 96.10% accuracy and 95.95% F1-score, outperforming HMM, CRF, BiLSTM, and IndicBERT-only baselines. The results demonstrate improved phrase boundary detection and BIO label consistency, although the model requires higher computational resources and has been evaluated on a single corpus. Future work will focus on cross-domain and multilingual evaluation, lightweight models, and integration with downstream NLP tasks.