Enhancing Neural Machine Translation of Low-Resource Languages: Corpus Development, Human Evaluation and Explainable AI Architectures

TL;DR

This work investigates transforming neural machine translation (NMT) for low-resource languages by combining careful Transformer hyperparameter optimisation, targeted in-domain corpus development, and rigorous human evaluation. The gaHealth health corpus for English↔Irish demonstrates substantial in-domain gains, while adaptNMT and adaptMLLM provide open-source environments for streamlined NMT development and fine-tuning of large multilingual language models, respectively. Across EN↔GA and EN↔MR pairs, Transformer architectures with well-chosen subword tokenisers (notably 16k BPE) and targeted domain data yield state-of-the-art results, with human MQM/SQM analyses confirming Transformer superiority over RNNs in accuracy and fluency. The work also foregrounds environmental sustainability (green reports) and presents a comprehensive roadmap for make MT more accessible, auditable, and greener through open-source tooling and explainable AI architectures.

Abstract

In the current machine translation (MT) landscape, the Transformer architecture stands out as the gold standard, especially for high-resource language pairs. This research delves into its efficacy for low-resource language pairs including both the English$\leftrightarrow$Irish and English$\leftrightarrow$Marathi language pairs. Notably, the study identifies the optimal hyperparameters and subword model type to significantly improve the translation quality of Transformer models for low-resource language pairs. The scarcity of parallel datasets for low-resource languages can hinder MT development. To address this, gaHealth was developed, the first bilingual corpus of health data for the Irish language. Focusing on the health domain, models developed using this in-domain dataset exhibited very significant improvements in BLEU score when compared with models from the LoResMT2021 Shared Task. A subsequent human evaluation using the multidimensional quality metrics error taxonomy showcased the superior performance of the Transformer system in reducing both accuracy and fluency errors compared to an RNN-based counterpart. Furthermore, this thesis introduces adaptNMT and adaptMLLM, two open-source applications streamlined for the development, fine-tuning, and deployment of neural machine translation models. These tools considerably simplify the setup and evaluation process, making MT more accessible to both developers and translators. Notably, adaptNMT, grounded in the OpenNMT ecosystem, promotes eco-friendly natural language processing research by highlighting the environmental footprint of model development. Fine-tuning of MLLMs by adaptMLLM demonstrated advancements in translation performance for two low-resource language pairs: English$\leftrightarrow$Irish and English$\leftrightarrow$Marathi, compared to baselines from the LoResMT2021 Shared Task.

Figures (38)

• Figure 1: Simplified representation of a feed-forward NN with 3 layers. A matrix of weights, i-w1 to i-w6 connects the input and hidden layers whereas a separate matrix h-w1 to h-w6 connects the hidden and output layers.
• Figure 2: Timeline of publications and mapping to thesis chapters. The illustration highlights the initial set of papers on automated machine learning (AutoML), followed by the MT papers and culminating in a book on the topic of MT and automation moorkens-lankford-way.
• Figure 3: Proposed approach of Transformers for low-resource languages
• Figure 4: BLEU performance for all model architectures
• Figure 5: TER performance for all model architectures