Bridging Dialects: Translating Standard Bangla to Regional Variants Using Neural Models
Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Bidyarthi Paul 1 , Tashreef Muhammad 1 · 2025
Read Bridging Dialects: Translating Standard Bangla to Regional Variants Using Neural Models by Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Bidyarthi Paul 1 , Tashreef Muhammad 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-05749.txt.
Introduction
Arafat Alam Khandaker and three fellow authors teach neural translators to speak regional Bangla, in a 2025 paper. Standard Bangla differs sharply from the dialects of Chittagong, Sylhet, Barishal, Noakhali, and Mymensingh in vocabulary, sound, and sentence shape, and those dialects are largely missing from language technology.
The team fine-tuned three translation models on a dataset of 32,500 sentences spanning the dialects. BanglaT5 performed best, with a character error rate of 12.3 percent and a word error rate of 15.7 percent, ahead of mT5 and mBART50. Dialect-aware tools could help speakers use services in their own variety of the language. The work also supports preserving linguistic diversity rather than flattening it into the standard form.
Catalog record
| CREATOR | Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Bidyarthi Paul 1 , Tashreef Muhammad 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-05749
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Bridging Dialects: Translating Standard Bangla to Regional Variants Using Neural Models Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Bidyarthi Paul 1 , Tashreef Muhammad 1 Abstract The Bangla language includes many regional dialects, adding to its cultural richness. The translation of Bangla Language into regional dialects presents a challenge due to significant variations in vocabulary, pronunciation, and sentence structure across regions like Chittagong, Sylhet, Barishal, Noakhali, and Mymensingh. These dialects, though vital to local identities, lack of representation in technological applications. This study addresses this gap by translating standard Bangla into these dialects using neural machine translation (NMT) models, including BanglaT5, mT5, and mBART50. The work is motivated by the need to preserve linguistic diversity and improve communication among dialect speakers. The models were fine-tuned using the “Vashantor” dataset, containing 32,500 sentences across various dialects, and evaluated through Character Error Rate (CER) and Word Error Rate (WER) metrics. BanglaT5 demonstrated superior performance with a CER of 12.3% and WER of 15.7%, highlighting its effectiveness in capturing dialectal nuances. The outcomes of this research contribute to the development of inclusive language technologies that support regional dialects and promote linguistic diversity. [email protected], [email protected], [email protected], [email protected] I Introduction Bangladesh is a country rich in linguistic diversity, with over 50 languages spoken across its 64 districts. Within the Bangla language itself, there are numerous regional dialects that vary significantly from one another in terms of vocabulary, pronunciation, tone, and sentence structure. These dialects, spoken in regions like Chittagong, Sylhet, Barishal, and Mymensingh, reflect the cultural and social uniqueness of each area. However, despite their prevalence, these dialects are often underrepresented in both academic research and technological applications. This study aims to bridge this gap by focusing on the translation of standard Bangla into these regional dialects using advanced neural machine translation (NMT) models. Neural Machine Translation (NMT) [ 1 ] is a method for translating words from one language to another. It is an innovative approach to machine translation that employs a single, massive neural network trained end-to-end and organized around an encoder-decoder architecture. The encoder converts a source sentence to a continuous space representation, and the decoder generates the translation from this representation. The impulsion for this research arises from the clear lack of studies dedicated to translating standard Bangla into its regional variants. Previous research has primarily concentrated on translating Bangla into other languages, such as English and Hindi, or on mixed languages like Banglish. Although in the study [ 2 ] they translated regional to Bangla language, there is a noticeable void in the reverse process—translating standard Bangla into its regional dialects. This gap is significant because it limits the understanding and accessibility of these dialects, which are an integral part of the cultural identity of their speakers. By addressing this gap, the study seeks to contribute to the preservation and promotion of linguistic diversity within Bangladesh. To tackle the challenge of translating standard Bangla into regional dialects, the study employs several cutting-edge neural models, including BanglaT5 [ 3 ] , mT5 [ 4 ] , and mBART50 [ 5 ] . These models are fine-tuned specifically to handle the linguistic complexities of the selected dialects. The approach involves training these models on a dataset that includes extensive samples of standard Bangla and its regional variants. The models leverage continuous representations and attention mechanisms, enabling them to focus on relevant parts of the input sentence and manage long-distance dependencies. By fine-tuning these models, the study aims to produce translations that are not only linguistically accurate but also culturally and contextually appropriate. The performance of these models is evaluated using established metrics like Character Error Rate (CER) [ 6 ] and Word Error Rate (WER) [ 7 ] . WER evaluates word-level accuracy by measuring the number of insertions, deletions, and substitutions in the translation, while CER captures more granular, character-level errors, critical for preserving linguistic details. The contribution of the research paper can be summarized as follows: • The research fills a significant gap in the field of translation studies by addressing the underexplored area of translating standard Bangla into regional dialects, a topic that has received limited attention in prior work. While earlier research has focused on translating Bangla into other languages or translating regional dialects into standard Bangla, this study uniquely contributes to the reverse process. • We fine-tuned several neural models, including BanglaT5, mT5, and mBART50, to translate standard Bengali into various regional dialects. We afterwards evaluated the performance of these models across different dialects to identify their strengths and areas for improvement. This study begins with an introduction in Section I , followed by a review of related works in Section II . The methodology is detailed in Section III , results are analyzed in Section IV , and limitations with future work are discussed in Section V . Finally, Section VI summarizes the study. II Related Works This section provides a compact overview of previous study which is relevant to our study, organized into three subsections: Transformer Based Neural Machine Translation, Adversarial Neural Machine Translation and Sequence-to-Sequence (Seq2Seq) Neural Machine Translation. II-A Transformer-Based Neural Machine Translation Fatema et al. [ 2 ] investigates the translation of Bangla regional dialects into standard Bangla using models like Bangla-bert-base, mT5 and BanglaT5. While BanglaT5 performs better in character error rate, word error rate, BLEU, and METEOR scores, Bangla-bert-base surpasses mBERT and mT5 in precision, recall, and F1-score . This work highlighted the superiority of specialized models in dialect translation over general-purpose models. Soran et al. [ 8 ] employs transformer-based NMT model for low-resource languages like Kurdish Sorani. It utilizes data from different sources and attention mechanisms . The model significantly enhances its translation performance and achieving a BLEU score of 0.45 for accurate translations. Laith et al. [ 9 ] proposes a novel Transformer-Based NMT model tailored for Arabic dialects, to solve low-resource language challenges like unique word order and limited vocabulary, resulting in higher BLEU scores. Significant advancement was demonstrated by extensive testing using several Arabic dialects and translations to Modern Standard Arabic (MSA). II-B Adversarial Neural Machine Translation Wenting et al. [ 10 ] tackles short sequence translation from Chinese to English using a generative adversarial network (GAN). The GAN has a generator that mimics human translations and a discriminator that tells them apart from real human translations. During training, dynamic discriminators and static BLEU score targets guide the generator. Tests on an English-Chinese dataset showed this method improved translation quality by over 8% compared to typical recurrent neural network (RNN) models, achieving an average BLEU score of 28.2. Lijun et al. [ 11 ] presentes Adversarial-NMT, a new neural machine translation (NMT) approach. Instead of just mimicking human translations, Adversarial-NMT uses a CNN adversary to minimize differences between human and NMT translations. The NMT model and adversary are co-trained using a policy gradient method. Tests on English-French and German-English tasks showed that Adversarial-NMT significantly improved translation quality over strong baseline models. Adversarial-NMT translated about 59.4% of sentences better than the baselines. II-C Sequence-to-Sequence (Seq2Seq) Neural Machine Translation The exploration of machine translation between English and Bangla using Recurrent Neural Networks (RNNs) is discussed in first two papers. Rafiqul et al. [