Bangla NER using Few-Shot Learning

Abstract

Named Entity Recognition (NER) is a critical task in Natural Language Pro- cessing (NLP) for low-resource languages like Bengali. This paper explores few-shot learning for Bengali NER using ProtoBERT, NN-SHOT, and Struct- SHOT models with mBERT, XLM-RoBERTa, and BanglaBERT embeddings. We addressed class imbalance in the BNER dataset through oversampling tech- niques, significantly enhancing model performance. In our 5-way 5-shot ex- periment, we observed that XLM-RoBERTa generally yielded the highest F1 scores: 0.4045 for ProtoBERT, 0.378 for NN-SHOT, and 0.3912 for StructSHOT. This study underscores the importance of balanced datasets and suggests future research on optimizing sampling strategies and advanced model architectures.