Improving Bangla Regional Dialect Detection Using BERT, LLMs, and XAI

Abstract

This research addresses the challenge of identifying and categorizing Bangla regional dialects through the application of sophisticated natural language processing techniques. Automated translation, digital content personalization, and speech recognition systems are all improved by precise dialect detection, which is a result of the linguistic diversity in Bangladesh. Few-shot learning techniques were used to compare the performance of transformer-based models, specifically Bangla BERT, with state-of-the-art large language models such as GPT-3.5 Turbo and Gemini 1.5 Pro, and to fine-tune them. The methodology entailed the utilization of a diverse regional Bangla speech samples from the Vasantor dataset, which encompassed regions including Mymensingh, Chittagong, Barishal, Noakhali, and Sylhet. In order to enhance the interpretability of the model, implementation of Local Interpretable Model-agnostic Explanations (LIME) was done. The Bangla BERT model achieved the highest accuracy of 88.74%. GPT-3.5 Turbo’s few-shot learning exhibited substantial potential, with an accuracy rate of 64%. These results underscore the significance of hyperparameter optimization and fine-tuning in the enhancement of regional dialect detection models.

Keywords

Citations by Year

Entered manually, not live-tracked.