Replay-Based Continual Adaptation for English-Norwegian Petroleum Machine Translation

Research
Machine Translation
LoRA / PEFT
Domain Adaptation
Evaluation
Replay-Based Continual Adaptation for English-Norwegian Petroleum Machine Translation

Tech Stack

Python
PyTorch
Transformers
Hugging Face
NLLB-200
LoRA
PEFT
SacreBLEU
chrF
Statistics

Description

This project studies this question in a source-update setting: an MT model is first adapted to an earlier NPD petroleum corpus and then updated with newer Equinor data. The goal is to compare different update strategies and understand the trade-off between learning from the newer source and retaining performance on the earlier source.

As the project and experimental lead, my contributions included:

  • Developing a continual-adaptation pipeline comparing single-source, mixed-source, sequential, and replay-based LoRA training.
  • Evaluating new-source adaptation, previous-source retention, and catastrophic forgetting using translation-quality and terminology metrics.

Research Question

How can an MT model learn from a new data source without forgetting what it learned from the previous source?

The answer is evaluated through a controlled comparison of single-source, mixed-source, continual, and replay-based update strategies.

Experimental Design

The experiments compare single-source training, mixed-source training, continual adaptation, and replay-based continual adaptation. All settings use fixed NPD and Equinor test sets. In total, 27 LoRA adapters were trained across three random seeds.

StrategyWhat it tests
NPD onlyPerformance before the model is updated with Equinor data.
Equinor onlyPerformance when the model is trained directly on Equinor data.
Mixed trainingPerformance when all NPD and Equinor data are used together.
Continual adaptationWhat happens when the NPD model is updated using only Equinor data.
Continual + replayWhether adding a small amount of NPD data during the Equinor update reduces forgetting.

Results

Update strategyNPD BLEUEquinor BLEUWhat it shows
NPD only61.1634.03Strong on NPD but weak on Equinor.
Equinor only50.3742.29Strong on Equinor but weak on NPD.
Mixed training59.9242.26Strong overall results using all data.
Continual adaptation54.7342.72Learns Equinor well but forgets some NPD knowledge.
Continual + 5% replay58.2142.11Reduces forgetting using a small amount of NPD data.
Continual + 10% replay59.2241.85Gives the best lightweight balance across both sources.

Main Finding

Replay-based continual adaptation gives the best practical update strategy. Continual adaptation improves performance on newer Equinor data but reduces performance on the earlier NPD source. Adding 5-10% NPD replay recovers much of the NPD loss while preserving most of the Equinor gain.

View Corpus Pipeline