Diagnosing Modular LoRA Experts for Multilingual Domain Translation

This project examines a modular adaptation hypothesis: if each source language has its own LoRA expert, can a learned router improve low-resource multilingual petroleum translation by selecting the right expert?

The study uses a frozen NLLB-200 backbone, language-specific LoRA experts, a learned router, and a synthetic corpus construction pipeline for German-, French-, and Dutch-to-Norwegian petroleum translation. The central finding is intentionally diagnostic: better routing accuracy does not necessarily translate into better BLEU or terminology accuracy.

My contribution was to frame the experiment as a routing diagnosis rather than a simple model-improvement story, implement the expert and router comparisons, analyze cross-expert transfer, and interpret why shared multitask adaptation and independent experts can outperform a learned-router MoE in this low-resource domain setting.

• Designed a modular LoRA expert setup with language-specific adapters and router-based expert selection.
• Developed the Target-Anchored Synthesis idea for creating multilingual petroleum-domain training pairs when parallel data is unavailable.
• Compared learned-router MoE, shared multitask LoRA, independent experts, and cross-expert transfer settings.
• Analyzed routing accuracy, BLEU, terminology accuracy, and expert specialization to separate routing quality from end-task translation quality.
• Presented the project conservatively as under review, with high-level methods and diagnostics but no public manuscript link.

Language Representation Diagnostics

Encoder representation analysis is used as one diagnostic lens for understanding whether multilingual source languages form separable routing signals.