Bridging Parametric Learning and Dynamic Non-Parametric Retrieval: A Verified Framework for Bangla Dialect Translation

Undergraduate Thesis (Course Code: ECE 452)
Department of Electronics and Communication Engineering
Hajee Mohammad Danesh Science and Technology University (HSTU), Dinajpur-5200, Bangladesh
October 2026

---

📌 Abstract & Overview

Bengali regional varieties like Chittagonian (~13M speakers) and Sylheti (~11M speakers) diverge sharply from Standard Colloquial Bengali (SCB) in phonology, morphosyntax, and vocabulary. Conventional Machine Translation (MT) encounters two major bottlenecks:

1. The Context-Pollution Dilemma in Seq2Seq: Naive multi-exemplar RAG collapses compact encoder-decoder models (e.g., NLLB drops to 3.02 BLEU) due to cross-attention disruption and positional embedding exhaustion.
2. Pure Parametric Misses: Standalone fine-tuning struggles on long-tail, uncodified dialectal lexemes and non-standard orthography.

This repository implements the Architecture-Aware Lean Adaptive RAG Framework, bridging parameter-efficient fine-tuning (LoRA on NLLB-200-Distilled-600M) with dynamic sparse-dense retrieval and relative hypothesis verification.

🌟 Key Highlights

• 🏆 State-of-the-Art Translation: 55.46 BLEU-4 overall (60.02 on Chittagonian, 51.14 on Sylheti) and 91.24% BERTScore-F1, outperforming 7B LLMs.
• ⚡ Ultra-Efficient Edge Footprint: Only 2.36M trainable parameters (0.38%); runs end-to-end within ~3.1 GB GPU VRAM (deployable on a single Tesla T4 or consumer RTX GPU).
• 🛡️ Zero-Leakage Benchmark: Rigorously audited 80:10:10 stratified split across 17,040 parallel pairs (0.0000% test-to-train data leakage verified via MinHash LSH and SHA-256 cluster hashing).
• 🔄 Safe Fallback Architecture: A competitive relative verifier enforces an empirical performance lower bound ($\delta \ge 0.005$), mathematically guaranteeing that retrieval never degrades parametric quality.

---

🏗️ System Architecture

flowchart TD
    A["Raw Dialect Input (Chittagonian / Sylheti)"] --> B["Deterministic Text Normalization\n(NFC, Nukta Unification, Khanda-Ta, Dari)"]
    B --> C{"Rare-Token Density Gating\n(D_rare >= 0.15?)"}
    
    %% Parametric Branch
    C -- "No (< 15%)\n[62% of traffic]" --> D["Parametric Path (k=0)\nDirect to LoRA-NLLB-600M"]
    D --> E["Generate Parametric Hypothesis (y_FT)"]
    
    %% Retrieval Branch
    C -- "Yes (>= 15%)\n[38% of traffic]" --> F["Dual-Index Retrieval Channel"]
    F --> G1["Sparse: BM25Okapi"]
    F --> G2["Dense: BGE-M3 (1024-dim)"]
    G1 & G2 --> H["Reciprocal Rank Fusion (RRF, k=60)"]
    H --> I["Extract Top-1 Lean Exemplar"]
    I --> J["Query-First Syntactic Formatting\nx + ' । প্রসঙ্গ: ' + y_tgt"]
    J --> K["Generate RAG Hypothesis (y_RAG)\nusing LoRA-NLLB-600M"]
    
    %% Verification & Arbitrator
    E & K --> L{"Competitive Relative Verifier\nSim(x, y_RAG) > Sim(x, y_FT) + 0.005\nAND |y_RAG| >= 0.5 |y_FT|"}
    L -- "Passed" --> M["Verified RAG Output (y_final)"]
    L -- "Failed / Drift" --> N["Parametric Fallback (y_FT)"]
    
    D --> O["Final Standard Bengali (SCB) Output"]
    M --> O
    N --> O

    style A fill:#e1f5fe,stroke:#0288d1,stroke-width:2px
    style O fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px
    style C fill:#fff9c4,stroke:#fbc02d,stroke-width:2px
    style L fill:#fce4ec,stroke:#d81b60,stroke-width:2px

Loading

---

🔬 Core Innovations

1. Deterministic Text Canonicalization

Resolves graphemic ambiguity across crowdsourced text and OCR scans:

• NFC Decomposition & Recomposition: Unifies visually identical byte streams.
• Nukta Unification: Maps decomposed consonants [ড, ঢ, য] + U+09BC into canonical precomposed glyphs (ড় U+09DC, ঢ় U+09DD, য় U+09DF).
• Word-Boundary Khanda-Ta: Maps ত + Hasanta (্) at word boundaries to ৎ (U+09CE).
• Bivariate Length Filtering: Enforces $0.33 \le r_{\text{word}}, r_{\text{char}} \le 3.00$ to eliminate truncation noise.

2. Dynamic Lexical Density Gating

Prevents prompt pollution on routine sentences: $$D_{\text{rare}}(x) = \frac{\sum_{w \in x} \mathbb{I}(w \in \mathcal{G}_{\text{dialect}})}{|x|}$$

• If $D_{\text{rare}}(x) < 0.15$: Routes directly through pure parametric generation (bypasses retrieval across ~62% of queries).
• If $D_{\text{rare}}(x) \ge 0.15$: Triggers lean single-exemplar retrieval.

3. Query-First Syntactic Formatting

Resolves positional embedding decay and query truncation in Seq2Seq decoders: $$\tilde{x} = x \circ \text{" । প্রসঙ্গ: "} \circ y_{\text{tgt}}$$ Positions the source dialect sentence $x$ at indices $0 \dots L_{\text{query}}-1$, keeping the encoder attention focused on the query while treating the exemplar as a soft modifier.

4. Competitive Relative Faithfulness Verification

Acts as a dynamic safety governor against distractor exemplars: $$\hat{y}{\text{final}} = \begin{cases} \hat{y}{\text{RAG}} & \text{if } \text{Sim}(x, \hat{y}{\text{RAG}}) > \text{Sim}(x, \hat{y}{\text{FT}}) + \delta ;\land; |\hat{y}{\text{RAG}}| \ge 0.5 \cdot |\hat{y}{\text{FT}}| \ \hat{y}{\text{FT}} & \text{otherwise} \end{cases}$$ Where $\delta = 0.005$. If retrieval introduces semantic drift or truncation, the system automatically falls back to $\hat{y}{\text{FT}}$.

---

📊 Experimental Results

Evaluated on the Kothon Test Benchmark ($N = 1,704$) across 10 evaluation metrics:

Model Architecture	Model Type	Parameters	VRAM	Chittagonian (BLEU-4)	Sylheti (BLEU-4)	Overall BLEU-4	chrF++	BERTScore-F1 (%)
BanglaT5 (Zero-Shot)	Seq2Seq	247M	~1.2 GB	1.91	1.90	1.91	8.40	60.94
M2M100-418M (Zero-Shot)	Seq2Seq	418M	~1.8 GB	13.37	9.52	11.31	23.42	76.02
NLLB-600M Base (Zero-Shot)	Seq2Seq	619M	~2.4 GB	26.92	21.27	23.94	36.45	81.93
Qwen2.5-7B (Zero-Shot)	Causal LLM	7.6B	~14.8 GB	28.53	21.58	24.98	40.00	82.87
Qwen2.5-7B (Hybrid-RAG, $k=4$)	Causal LLM	7.6B	~14.8 GB	50.77	34.08	41.83	56.22	86.76
NLLB-600M (Static RAG)	Seq2Seq	619M	~2.4 GB	3.27	2.81	3.02	19.62	61.44
Proposed FT-NLLB + Adaptive RAG + Verifier	Seq2Seq + LoRA	619M (2.36M trainable)	~3.1 GB	60.02	51.14	55.46	61.46	91.24

Key Takeaway

• Outperforms zero-shot NLLB-600M by +31.52 BLEU-4 (+131.7% relative).
• Outperforms 7B Qwen2.5-7B Hybrid-RAG by +13.63 BLEU-4 while requiring 79% less VRAM (~3.1 GB vs ~14.8 GB).

---

📂 Repository Structure

├── configs/
│   ├── lora_config.json            # LoRA hyperparameters (r=16, alpha=32)
│   └── training_args.yaml          # FP32/BF16 training protocol config
├── data/
│   ├── kothon_splits/              # Leak-free 80:10:10 dataset splits
│   └── gloss_dictionaries/         # Top 2,000 statistical bilingual lexicons
├── src/
│   ├── canonicalization/
│   │   ├── normalizer.py           # Unicode NFC, Nukta, Khanda-Ta canonicalization
│   │   └── guardrails.py           # Bivariate length ratio (0.33 - 3.00) filtering
│   ├── retrieval/
│   │   ├── bm25_indexer.py         # BM25Okapi sparse retrieval
│   │   ├── dense_indexer.py        # BGE-M3 1024-dimensional dense indexing
│   │   └── rrf_fusion.py           # Reciprocal Rank Fusion (k=60)
│   ├── models/
│   │   ├── gating.py               # Lexical density gating (threshold = 0.15)
│   │   ├── modeling_nllb_lora.py   # LoRA-adapted NLLB-200 architecture
│   │   └── verifier.py             # Competitive relative faithfulness verifier
│   └── pipeline.py                 # End-to-end verified inference pipeline
├── scripts/
│   ├── train_lora.py               # Training script with FP32 Master / BF16
│   ├── build_indices.py            # FAISS + BM25 index generator
│   └── evaluate_metrics.py         # 10-metric evaluation suite
├── requirements.txt
└── README.md

---

🚀 Quickstart & Usage

1. Installation

git clone https://github.com/YOUR_USERNAME/bangla-dialect-translation.git
cd bangla-dialect-translation

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

2. Quick Inference Example

from src.pipeline import VerifiedDialectTranslator

# Initialize translator (runs on a single GPU or CPU)
translator = VerifiedDialectTranslator.from_pretrained(
    base_model="facebook/nllb-200-distilled-600M",
    lora_weights="checkpoints/lora-nllb-kothon",
    bm25_index="data/indices/bm25.pkl",
    dense_index="data/indices/bge_m3_faiss.index",
    device="cuda"
)

# Translate Chittagonian
text_ctg = "খাল্া, আঁই এহতনা ন জানি ।"
result = translator.translate(text_ctg, dialect="chittagonian")
print(result["translation"])
# Output: "খালা, আমি এখনো জানি না।"
print(f"Path taken: {result['path']}, Verification Margin: {result['margin']}")

# Translate Sylheti
text_syl = "আমি তো লাফও দিতাম পারি।"
result = translator.translate(text_syl, dialect="sylheti")
print(result["translation"])
# Output: "আমি তো লাফও দিতে পারি।"

3. Training LoRA Adapters

python scripts/train_lora.py \
    --base_model facebook/nllb-200-distilled-600M \
    --train_data data/kothon_splits/train.jsonl \
    --val_data data/kothon_splits/val.jsonl \
    --output_dir checkpoints/lora-nllb-kothon \
    --lora_rank 16 \
    --lora_alpha 32 \
    --fp32_master_bf16 True \
    --batch_size 4 \
    --gradient_accumulation_steps 8 \
    --epochs 4

---

---

👥 Authors & Acknowledgments

• Undergraduate Researchers:
  • Student ID: 2102112
  • Student ID: 2102129
  • Student ID: 2102155
  • Session: 2021
• Affiliation: Department of Electronics and Communication Engineering (ECE), Hajee Mohammad Danesh Science and Technology University (HSTU).
• Gratitude: We dedicate this work to our beloved parents and acknowledge the creators of the Kothon benchmark and open-source models (Meta AI's NLLB, BAAI's BGE-M3).