Parameter-Efficient Neural Machine Translation for Low-Resource Languages

This research project investigates parameter-efficient adaptation of neural machine translation for a low-resource English-Norwegian petroleum domain. The work fine-tunes NLLB-200 distilled 600M with LoRA adapters and evaluates how domain-specific adaptation behaves under constrained data and compute settings.

The study combines data-quality assessment, corpus cleaning, LoRA hyperparameter search, model-scale evaluation, and comparison against full fine-tuning and commercial machine translation systems. The resulting paper, LoRA Fine-Tuning of English-Norwegian NMT for the Oil & Gas Industry, was published at EAMT.

My contribution covered the end-to-end experimental pipeline: data processing, training orchestration, LoRA configuration search, evaluation with BLEU and chrF, analysis of adaptation trade-offs, and preparation of reproducible code and documentation.

• Built a reproducible English-Norwegian petroleum-domain NMT pipeline using NLLB-200, LoRA, and Hugging Face tooling.
• Designed corpus diagnostics for alignment quality, text length distributions, vocabulary coverage, and domain terminology patterns.
• Optimized LoRA configurations with grid search and Optuna/ASHA, identifying key hyperparameter effects on translation quality.
• Compared parameter-efficient LoRA adaptation with full fine-tuning and external MT baselines under low-resource conditions.
• Evaluated translation quality with BLEU, chrF, and terminology-oriented analysis for domain-specific adequacy.

Corpus Diagnostics

Exploratory analysis of the English-Norwegian petroleum corpus, including source and target length distributions, length-ratio filtering, and source-target alignment behavior.

Domain Vocabulary Analysis

Frequency analysis highlighting petroleum-domain terminology and distributional patterns in the English source and Norwegian target data.