Research series

LoRA for Low-Resource Domain NMT

A research-facing path through low-resource domain MT, LoRA adaptation, data quality, HPO, terminology evaluation, data scaling, and the project story behind English--Norwegian petroleum translation.

Start the series
中文
Multilingual AI

Part I

Research problem

  1. 1. Why Low-Resource Domain MT Is Hard
  2. 2. Why LoRA Fits Low-Resource Domain Adaptation

Part II

Method and adaptation

  1. 1. LoRA for Neural Machine Translation
  2. 2. Hyperparameter Optimization for LoRA NMT
  3. 3. LoRA vs Full Fine-Tuning: What Do We Actually Save?

Part III

Data-centric MT

  1. 1. Data Quality Diagnostics for Parallel Corpora
  2. 2. Terminology as an Evaluation Problem in Domain MT
  3. 3. Data Scaling for LoRA Domain Adaptation

Part IV

Evaluation and research story

  1. 1. How to Evaluate Machine Translation Beyond BLEU
  2. 2. My LoRA NMT Project: Research Question, Method, Findings, and Limitations

Why this belongs in Research, not Foundations

The mathematical ideas are present, but they serve the research claim: why this method, this data, and this evaluation protocol make sense.

Problem-driven

The series starts from low-resource domain MT and technical risk, not from abstract linear algebra.

Method-aware

LoRA is explained through adaptation capacity, controlled updates, HPO, and full fine-tuning trade-offs.

Evaluation-focused

BLEU, chrF, COMET, terminology metrics, and human error analysis are treated as evidence design.

The series ends with a thesis-style research narrative.

It is designed to help explain the project clearly in interviews and future PhD conversations.

Read the project story
中文版本