Domain-Specific Vietnamese-Bahnaric Translation

• The paper demonstrates robust data augmentation strategies (MTL DA with Token+Swap and SBA) that boost BLEU scores by up to 11.44 points over the baseline.
• It details the use of Transformer-based architectures with syllable-level tokenization, beam search, and fine-tuning using AdamW in low-resource settings.
• The study provides practical guidelines for domain-specific translation by oversampling synthetic data while minimizing reliance on external language resources.

Domain-specific Vietnamese-Bahnaric translation refers to the application of neural machine translation (NMT) techniques to produce high-quality translations between Vietnamese and Bahnaric within specialized domains (e.g., government, education, narrative texts, or conversational language), especially under conditions of limited parallel corpora and language resources. This challenge is amplified by the under-resourced status of Bahnaric and the need to preserve specialized terminology and stylistic nuance. Recent research converges on data augmentation, Transformer-based architectures, and domain adaptation strategies as the primary avenues for progress.

1. Dataset Characteristics and Preprocessing

The parallel corpora available for Vietnamese-Bahnaric translation are considerably smaller than those for high-resource language pairs. Nguyen et al. (Nguyen et al., 27 Jan 2026) report a training set of 16,105 sentence pairs, with validation and test sets of 1,987 and 1,988 pairs, respectively. The domains span formal and informal conversational sentences, formal greetings, narrative stories, folk tales, and governmental/educational text. Preprocessing includes cleaning stray symbols and syllable-level tokenization for both languages, reflecting the orthographic structures (e.g., Bahnar-Kriem orthography). No external monolingual data, back-translation, or bilingual dictionaries are incorporated, except for alignment during specific augmentation routines.

2. Augmentation Strategies for Low-Resource Scenarios

Two primary sentence-oriented data augmentation (DA) schemes have been demonstrated to enhance NMT performance on Vietnamese-Bahnaric (Nguyen et al., 27 Jan 2026):

• Multi-Task Learning Data Augmentation (MTL DA): Five auxiliary tasks apply radical noising to the target (Bahnaric) side:
  1. Swap: Randomly exchange a fraction (α=0.5 by default) of tokens with others in the same sentence.
  2. Token masking: Replace α · |y| random tokens with the UNK token.
  3. Reverse: Reverse the target sentence word order.
  4. Source copying: Replace the target with a verbatim copy of the source (yields negative results).
  5. Replace: Substitute α · |y| source-target aligned token pairs with random vocabulary.

  Synthetic datasets from these transformations are appended to the parallel corpus, effectively enlarging its size without requiring extra data. The “Token+Swap” method yielded the highest BLEU increase (+10.75 over baseline, 40.64 vs. 29.89).

• Sentence Boundary Augmentation (SBA): This simulates imperfect segmentation typical in oral or historical texts by randomly truncating and splicing adjacent sentence pairs. Algorithmically:
  • Randomly choose a truncation fraction p (0.1–0.9; p=0.7 yielded best BLEU).
  • Merge rear and leading sub-segments from consecutive pairs to create synthetic sentence pairs.
  • SBA resulted in a BLEU score of 41.33 (+11.44 over baseline).

These methods require no external monolingual data or auxiliary systems, making them highly suited for truly low-resource, domain-specific settings (Nguyen et al., 27 Jan 2026).

3. Model Architectures and Training Protocols

All state-of-the-art regimes for Vietnamese-Bahnaric domain-specific translation deploy Transformer-based sequence-to-sequence models. The standard backbone is a pre-trained model such as BARTpho_syllable (6 encoder and 6 decoder layers, d_model=768, n_heads=12, d_ff=3072, dropout=0.1) (Nguyen et al., 27 Jan 2026). Training is performed using the HuggingFace Transformers library with AdamW, a constant learning rate of 2e-5, batch size 32, and typically two epochs. Beam search (beam=5) is adopted for decoding. Only the cross-entropy loss over the autoregressive decoder is computed: $$\mathcal{L}(\theta) = -\sum_{(x,y)\in\mathcal{D}} \sum_{t=1}^{T} \log P_\theta(y_t \mid y_{<t}, x)$$ No modifications to the model architecture are required for DA integration.

For more general (non-Bahnaric) low-resource scenarios, other works detail leveraging larger “out-of-domain” corpora, synthetic parallel data via back-translation, and adversarial shared/private encoder-decoder splits (Gu et al., 2019, Moslem et al., 2022). These can be considered for Vietnamese-Bahnaric as well, especially when comparable data is available.

4. Domain Adaptation and Synthetic Data Generation

Where in-domain parallel data are scarce, advanced domain adaptation protocols from general NMT have been shown to generalize:

• Separation of Domain-Invariant and Domain-Specific Features: Implemented by maintaining a shared encoder/decoder for invariant features and private encoder/decoders per domain, with adversarial minimax (via a domain discriminator) to enforce domain-agnostic representations. This architecture allows out-of-domain data to be exploited during in-domain fine-tuning, which is especially relevant when comparable data (e.g., religious texts, government bulletins) is available (Gu et al., 2019).
• Synthetic Data via Pre-trained LLMs and Back-translation: Synthetic in-domain parallel data can be generated using either a small in-domain parallel “seed” or only in-domain monolingual source. The process comprises:
  1. Target-side generation using a pre-trained LM (multilingual if Bahnaric LM is unavailable), sampling continuations conditioned on in-domain prompts.
  2. Back-translation using the baseline NMT model to generate pseudo-source sentences.
  3. Mixed fine-tuning with aggressive oversampling of synthetic data (mixture weight α=0.9 in practice).

Reported gains in several language pairs range from +2 to +6 BLEU, while preserving overall translation quality. There is emphasis on prompt engineering (adding domain tags), sampling strategies (top-k, top-p), tagging synthetic data, and maintaining a small in-domain validation/test set for robust evaluation (Moslem et al., 2022).

5. Empirical Performance and Evaluation

The Nguyen et al. experiments (Nguyen et al., 27 Jan 2026) benchmark DA methods on both overall BLEU and error-specific categories (collocation and word-by-word translations):

Method	BLEU	Δ vs. Baseline
Baseline	29.89	—
Easy Data Augment	36.37	+6.48
Semantic Embedding	39.20	+9.31
MTL DA (Token+Swap)	40.64	+10.75
Sentence Boundary (p=0.7)	41.33	+11.44

In error-focused ablations (collocation and word-by-word issues in sentences with per-sentence BLEU 0.2–0.4), both MTL DA and SBA outperformed the baseline, with SBA providing the best aggregate improvement. No human evaluation, chrF, or TER metrics are reported in the Bahnaric literature to date, but general NMT adaptation studies report consistent gains in automatic and human metrics (Moslem et al., 2022).

6. Practical Guidelines and Recommendations

For domain-specific Vietnamese-Bahnaric MT in low-resource settings:

• Data augmentation is essential: Adopt both MTL DA (preferably “Token+Swap” noise) and SBA, which have complementary effects and yield maximal BLEU improvements (Nguyen et al., 27 Jan 2026).

• Apply syllable-level tokenization: Given the orthographic characteristics of both languages.
• Leverage pre-trained Vietnamese or multilingual models: Fine-tune these backbones with the DA-augmented corpus.
• Oversample in-domain and synthetic data for fine-tuning: Maintain a 9:1 ratio in favor of domain-augmented data per step (Moslem et al., 2022).
• Minimal reliance on external resources: Both leading Bahnaric strategies report results without additional monolingual data or large dictionaries. When such resources become available, back-translation and domain-invariant/private modeling can be introduced (Gu et al., 2019).
• Rigorous validation: Hold out 50–200 in-domain sentence pairs for spBLEU or BLEU evaluation, focusing on terminology consistency and translation acceptability.
• Phrase and terminology extraction: Use hard lexical constraints at test time if preservation of specific terms is critical (Moslem et al., 2022).

7. Future Directions and Limitations

Current state-of-the-art DA methods for Vietnamese-Bahnaric translation do not require complex preprocessing or extra systems, streamlining their deployment in real-world settings (Nguyen et al., 27 Jan 2026). However, the literature acknowledges that their effectiveness has not yet been tested in hybrid pipelines combining sentence-level DA with back-translation, nor with larger Transformer models. No systematic human evaluation has been reported for the Bahnaric context. Expanding to other dialects, integrating external monolingual data, or leveraging the shared/private adversarial NMT frameworks are plausible next steps, as are adaptation to extremely low-resource settings and broader domain adaptation studies (Gu et al., 2019, Moslem et al., 2022, Nguyen et al., 27 Jan 2026).