Serving the Underserved: Leveraging BARTBahnar Language Model for Bahnaric-Vietnamese Translation

Long S. T. Nguyen, Tran T. B. Le, Huong P. N. Nguyen, Quynh T. N. Vo, Phong H. N. Nguyen, Tho T. QuanNAACL 2025

TL;DR

This paper tackles the challenge of building a machine translation system for Bahnaric, a low-resource language spoken by an ethnic minority in Vietnam. They introduce BARTBahnar, a novel hybrid system that leverages transfer learning from a pre-trained Vietnamese model and incorporates rule-based methods to achieve significant improvements in Bahnaric-Vietnamese translation accuracy, crucial for cultural preservation and communication.

What Problem Does This Paper Address?

Machine translation (NMT) has seen remarkable progress, but its effectiveness is heavily dependent on large, parallel datasets. This is a major hurdle for "low-resource" languages, where such data is scarce. The Bahnar people, an ethnic minority in Vietnam, face this challenge. They have a rich cultural heritage, but limited access to modern technologies, and a lack of effective tools for translating between their language and Vietnamese hinders cultural exchange and preservation. Existing NMT architectures, particularly encoder-decoder models, are well-suited for translation, but building them for languages like Bahnaric is difficult due to data scarcity, limited vocabulary, and the absence of parallel corpora. This paper aims to bridge this gap by developing an effective Bahnaric-Vietnamese translation system.

Methods

The core of the proposed solution is a novel hybrid NMT architecture. The system's language model, BARTBahnar, is built upon the BARTPho model, a pre-trained Vietnamese sequence-to-sequence model. The adaptation process involves:

  1. Continual Pre-training on Vietnamese: Leveraging BARTPho's existing knowledge of Vietnamese grammar and syntax.
  2. Continual Pre-training on Bahnaric: Adapting BARTPho to Bahnaric by training it on a monolingual Bahnaric corpus using an autoregressive Masked Language Modeling (MLM) objective. This stage involved extensive field surveys to gather linguistic materials.
  3. Fine-tuning on Bilingual Data: The model is then fine-tuned on a bilingual Bahnaric-Vietnamese dataset.

To address data scarcity, various Data Augmentation (DA) techniques are employed, including:

  • Swapping: Reordering sentence segments.
  • Combining: Merging semantically related sentences.
  • Replacing: Substituting words with synonyms or thematic labels.
  • Insertion/Deletion: Adding or removing non-essential words.
  • Sliding Window: Extracting overlapping sub-sequences.

Beyond the language model, the system incorporates a hybrid approach with rule-based and statistical methods for:

  • Loanword Detection: Identifying and excluding shared loanwords between Bahnaric and Vietnamese.
  • Word Segmentation: Splitting Bahnaric sentences into meaningful phrases using Pointwise Mutual Information (PMI).
  • Lexical Mapping: Using a bilingual dictionary and Solr for efficient mapping of phrases to Vietnamese equivalents.
  • Post-Processing: Standardizing punctuation, capitalization, and word order, and resolving ambiguities using a scoring mechanism based on a pre-trained language model.

Datasets

The paper introduces a newly collected bilingual Bahnaric-Vietnamese dataset, referred to as the "Original dataset." This dataset was compiled through:

  • Direct interviews with native Bahnar speakers.
  • Analysis of printed texts (religious books, newspapers, song lyrics).
  • Collection of local news bulletins and historical documents.

Additionally, a larger dataset was created using back-translation techniques, reconstructing synthetic Bahnaric text from Vietnamese Wikipedia. The statistics for these datasets are:

  • Original: 53,942 sentence pairs.
  • Back-Translation: 270,587 sentence pairs.
  • Total: 324,529 sentence pairs.

For training and evaluation, the "Original" dataset was split, with 90% for training and 10% for testing.

Key Insights and Ideas

  • Transfer Learning is Crucial for Low-Resource Languages: The success of BARTBahnar highlights the power of leveraging pre-trained models from related, higher-resource languages (Vietnamese in this case) as a foundation. This significantly reduces training costs and data requirements.
  • Domain Adaptation is Non-Negotiable: Simply using a pre-trained Vietnamese model (BARTPho) without specific adaptation to Bahnaric yields suboptimal results. Continual pre-training on Bahnaric data is essential for capturing its unique linguistic features.
  • Hybrid Approaches Offer Robustness: Combining a powerful neural model (BARTBahnar) with rule-based methods (loanword detection, lexical mapping) and statistical techniques (word segmentation) provides a more robust translation system, especially for handling specific linguistic phenomena like loanwords and ambiguous phrases.
  • Careful Data Augmentation is Key: Not all DA techniques are beneficial. While some, like synonym replacement, significantly boost performance, others can introduce noise and degrade quality if applied indiscriminately. The paper emphasizes the importance of selecting DA strategies that preserve semantic fidelity and grammatical correctness.
  • Linguistic Similarities are a Valuable Asset: The shared linguistic background and the presence of loanwords between Bahnaric and Vietnamese are identified as critical insights that inform the design of the translation system.

Key Results

The paper reports substantial improvements in Bahnaric-Vietnamese translation quality.

  • BARTBahnar consistently outperforms all baselines, including Transformer, PhoBERT-Fused NMT, ViT5, and BARTPho (without Bahnaric adaptation).
  • On the Bahnaric-Vietnamese bilingual dataset, BARTBahnar achieved a BLEU score of 10.41 and a METEOR score of 0.2822. This is a significant improvement over the next best baseline, BARTPho (5.73 BLEU, 0.2076 METEOR).
  • The analysis of Data Augmentation methods showed that the "Replace (synonym)" method yielded the greatest improvement, increasing translation accuracy by up to 200% in certain configurations.

Limitations and Future Work

The authors acknowledge several limitations:

  • Reliance on a "Sibling" Language: The transfer learning approach relies on the availability of a related, higher-resource language (Vietnamese). This might not be feasible for extremely low-resource languages lacking such a counterpart. Training a model from scratch would be prohibitively expensive.
  • Structural Similarities Dependency: The effectiveness of the transfer learning hinges on structural similarities between languages. Adapting this approach to languages with drastically different syntax could be challenging.
  • Bilingual Dictionary Requirement: The rule-based components necessitate a bilingual dictionary, which can be labor-intensive to develop and may suffer from coverage issues if the source corpus is limited.

Future work could involve:

  • Further customizing the language model by integrating more Bahnaric-specific linguistic properties.
  • Refining rule-based components to handle more nuanced text.
  • Exploring additional combinations of data augmentation methods.

Summary updated

BibTeX
@inproceedings{nguyen2025serving,
  author = {Long S. T. Nguyen and Tran T. B. Le and Huong P. N. Nguyen and Quynh T. N. Vo and Phong H. N. Nguyen and Tho T. Quan},
  title = {{Serving the Underserved: Leveraging BARTBahnar Language Model for Bahnaric-Vietnamese Translation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure