Research map / Large language models / Language, safety and interpretability
Revitalizing Bahnaric Language through Neural Machine Translation: Challenges, Strategies, and Promising Outcomes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds BN-BARTpho, a BARTpho-based Vietnamese-to-Bahnar NMT system with a chunk/anchor pipeline and multi-task-style data augmentation, reporting BLEU gains over Transformer and PhoBERT-fused baselines.
Problem
Vietnamese-to-Bahnar machine translation for an extremely low-resource ethnic minority language with scarce parallel data, small vocabulary, and no fine-tuning resources.
Approach
Fine-tune a Vietnamese pretrained seq2seq model (BARTpho) on a Bahnar parallel corpus with frozen encoder, combine it with dictionary mapping for anchors and entities, and apply multi-task-style word-level and sentence-boundary augmentation adapted for NMT.
Key results
- BN-BARTpho reaches 49.20 BLEU, above Transformer (28.78), PhoBERT-fused NMT (38.49), Loanformer (42.63), and BARTphoEncoderPGN (47.91), and slightly above PE-PD-PGN (49.00).
- Sentence boundary augmentation alone raises BLEU to 54.31 over the 49.20 baseline; swap+token gives 53.91.
- Source-copy augmentation collapses BLEU to 2.59, while reverse drops it to 39.79.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BN-BARTpho instead of the strongest baseline PE-PD-PGN | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.00 -> 49.20 | Helped |
| Add swap augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 52.94 | Helped |
| Add token augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 51.80 | Helped |
| Add source-sentence augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 2.59 | Hurt |
| Add reverse augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 39.79 | Hurt |
| Add replace augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 51.33 | Helped |
| Add swap and replace augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 51.57 | Helped |
| Add swap and token augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 53.91 | Helped |
| Add replace, token, and swap augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 52.89 | Helped |
| Add sentence boundary augmentation (ablation) | Vietnamese-Bahnar translation · Vietnamese-Bahnar | BLEU: 49.20 -> 54.31 | Helped |
Surprising findings
- Copying the source sentence as an augmentation target drastically hurts performance (BLEU 2.59).
- Combining all three best augmentation tasks (swap, token, replace) did not beat the swap+token pair.
Limitations
- Evaluation relies only on BLEU on a small private dataset collected from one region (Binh Dinh).
- Dataset is not released, limiting reproducibility.
- Augmentation hyperparameter alpha was fixed at 0.5 with limited exploration reported; no human evaluation.
Experimental setup
Datasets: Vietnamese-Bahnar parallel corpus (Binh Dinh dialect) Vietnamese-Bahnar bilingual dictionary
Benchmarks: Vietnamese-Bahnar bilingual test set (1,988 pairs)
Metrics: BLEU
Baselines: Transformer PhoBERT-fused NMT Loanformer BARTphoEncoderPGN PE-PD-PGN
Builds on: BARTpho BERT-fused NMT Multi-task learning data augmentation (Sánchez-Cartagena et al. 2021) EDA Sentence Boundary Augmentation
Closest papers by meaning
- ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation (AAAI 2026 · similarity 0.87)
- Neural Machine Translation with Joint Representation (AAAI 2020 · similarity 0.86)
- Evaluating the Cross-Lingual Effectiveness of Massively Multilingual Neural Machine Translation (AAAI 2020 · similarity 0.86)
- Rephrasing the Reference for Non-autoregressive Machine Translation (AAAI 2023 · similarity 0.85)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.85)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.