Recent advances in Bangla NLP: a survey of few-shot, zero-shot, and transfer learning techniques

Abstract

Bangla gets spoken by a ton of people, but in natural language processing, it's seen as a low resource language. That is mostly because there is not much annotated data out there, plus the grammar can be really tricky, and benchmarks for specific tasks are kind of lacking. This paper looks at how modern ways of learning and pretrained models work for Bangla NLP, doing a bunch of empirical stuff to check them out. They go through transfer learning, few shot, and zero shot methods, using models like mBERT, XLM R, mT5, BLOOM, those GPT style ones, and even BanglaBERT. All this on tasks such as sentiment analysis, named entity recognition, machine translation, question answering, and summarization. The experiments show that transfer learning, especially with models adapted for Bangla or multilingual ones, tends to do better than few shot or zero shot in terms of being robust and accurate. But I think prompt-based approaches still hold up pretty well in situations where resources are super low. It seems like how you preprocess the language matters a lot, and picking the right metrics for evaluation can change things too. Our findings, well, give some practical ideas for making Bangla NLP systems that scale and include more people. Not everything is fully sorted out here, like how to balance all these factors perfectly.