Modular Expert Architectures for Multilingual Domain Adaptation : Parameter-Efficient Norwegian Petroleum Translation with LoRA and Gated Routing

Sammanfattning

High-stakes industrial domains such as petroleum engineering require highlyaccurate terminology translation, yet multilingual neural machine translationsystems remain unreliable in low-resource, domain-specific settings. This chal-lenge is particularly acute in Norwegian petroleum translation due to scarceparallel corpora, cross-lingual interference in multilingual fine-tuning, and stricton-premise deployment constraints.We propose a parameter-efficient multilingual adaptation framework built ona frozen NLLB-200-distilled-600M backbone, combining language-specific Low-Rank Adaptation (LoRA) experts with a learned gated routing mechanism. Tosupport low-resource language pairs beyond English–Norwegian, we introduceTarget-Anchored Synthesis, a synthetic data generation pipeline that producespetroleum-domain parallel data for German, French, and Dutch. The pipeline isanchored on validated English–Norwegian pairs and filtered using LaBSE simi-larity, Formal Terminology Accuracy, and back-translation consistency, yieldingapproximately 43,000 sentence pairs.Experimental results show that synthetic-data experts achieve 93–96% ofEnglish–Norwegian BLEU performance across additional language pairs, and im-prove terminology accuracy in Dutch–Norwegian translation by 11.7 percentagepoints over Google Translate. A consistent trade-off emerges: shared-parameteradaptation improves surface fluency and BLEU, while language-specific expertsachieve higher terminology fidelity.Ablation and routing analyses reveal systematic confusion between closelyrelated languages, especially Dutch and German. However, improving routingaccuracy yields limited downstream gains, and cross-expert evaluations showthat non-selected experts remain competitive. These results indicate that routingaccuracy is not the primary bottleneck in modular multilingual adaptation; in-stead, performance is strongly shaped by partially shared target-side translationcapacity across language-specific experts.Overall, this thesis presents a parameter-efficient and terminology-awaremultilingual adaptation framework for industrial translation, and shows thatimproving routing or expert specialization alone is insufficient without addressingshared target-language capacity in low-resource domain settings.