JoshMcGiff/SemiAdapt

[LREC 2026] Repository for the paper: "SemiAdapt: Semi-Supervised and Efficient LoRA-Based Domain Adaptation for Low-Resource Irish Machine Translation with Transformers" # SemiAdapt: Semi-Supervised and Efficient LoRA-Based Domain Adaptation for Low-Resource Irish Machine Translation with Transformers --- ## 🧩 Overview This repository contains code and model config for **SemiAdapt-Full** and **SemiAdapt-LoRA**, two efficient domain adaptation methods for Transformer-based Neural Machine Translation (NMT). Our approaches target **low-resource language translation**, with a case study on **Irish**. Traditional full-model fine-tuning of large multilingual models (billions of parameters) is computationally expensive. **SemiAdapt-Full** and **SemiAdapt-LoRA** address this challenge by improving **parameter-efficient fine-tuning (PEFT)** techniques such as LoRA. Our methods achieve strong adaptation while reducing memory and training cost. --- ## 🚀 Key Contributions - **SemiAdapt-Full**: A fine-tuning approach that enhances domain adaptation efficiency and can outperform full fine-tuning. - **SemiAdapt-LoRA**: A parameter-efficient variant that matches or exceeds full-model fine-tuning performance. - **Embedding-based inference** for improved performance on large, noisy corpora and efficient inference overall. - Comprehensive experiments on **Irish translation**. ---