AgriBanglaT5 — Agricultural English→Bengali Translation

A LoRA adapter for csebuetnlp/banglat5_nmt_en_bn that specialises it for agricultural advisory translation (English → Bengali). Stock BanglaT5 mistranslates agricultural terminology (pesticide / disease / crop names); this adapter, trained on the AgriEnBn corpus with glossary term hints, renders the domain vocabulary correctly and translates more fluently — at the same 247M size and CPU latency.

Results (AgriEnBn gold test set, 198 human-verified sentences)

Metric	Stock BanglaT5	AgriBanglaT5
chrF++	48.1	54.2 (+6.1)
BLEU	23.4	28.1 (+4.7)
TermAcc (agri terms)	76.4%	81.8% (+5.4 pp)
Negation faithfulness	89.7%	93.1% (+3.4 pp)
NumberMatch	97.2%	96.3% (tie)

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from peft import PeftModel

base = "csebuetnlp/banglat5_nmt_en_bn"
tok = AutoTokenizer.from_pretrained(base, use_fast=False)
model = PeftModel.from_pretrained(AutoModelForSeq2SeqLM.from_pretrained(base),
                                  "Sabuktagin/AgriBanglaT5").merge_and_unload().eval()

# For the full terminology gain, append the canonical Bengali term as a hint
# after each glossary term in the source, e.g.:
#   "Apply Tricyclazole ⟦ট্রাইসাইক্লাজল⟧ at 0.6 gram per liter."
# (see the AgriEnBn repo for the glossary + helper).
src = "Spray Imidacloprid ⟦ইমিডাক্লোপ্রিড⟧ against brown planthopper ⟦বাদামি গাছফড়িং⟧."
ids = tok(src, return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, num_beams=5, max_length=192)[0], skip_special_tokens=True))

Training

• Data: AgriEnBn — 3,034 EN→BN pairs (back-translation of recovered Bijoy-encoded Bangladeshi agronomy documents + EN-side distillation), LaBSE-filtered ≥0.70.
• Method: LoRA (r=16, α=32) on the q,v projections; glossary inline term hints on the source. 3 epochs; held-out eval loss 2.56→2.36.

Limitations

• Specialist for agriculture (esp. rice); not intended for general-domain text.
• Rare numeric errors remain (one pH range in the test set). Verify dosages.
• Trained partly on back-translated (synthetic) data.
• Non-commercial (CC BY-NC-SA 4.0), inherited from BanglaT5 / BanglaNMT.

Citation

Cite this work, plus BanglaT5 (Bhattacharjee et al., EACL Findings 2023) and BanglaNMT (Hasan et al., EMNLP 2020).