The Open University
Developing Machine Translation for bioinformatics: An exploration into domain-specific terminology, domain adaptation, and evaluation.
Abstract
dc:description.abstractMachine Translation (MT) can support knowledge exchange in specialized scientific domains. While prior research often focused on areas such as biomedical sciences, the closely related field of bioinformatics remains underexplored. As bioinformatics becomes central to the Life Sciences, the demand for multilingual content and domain-adapted MT resources grows. Bioinformatics also poses distinctive challenges, including dense specialized terminology, including Multiword Expressions (MWEs), and scarce parallel data. This thesis investigates the linguistic profile of bioinformatics and how MT systems can be adapted and evaluated for English–French translation. Dedicated monolingual and parallel corpora were compiled, and synthetic parallel data were created via paraphrasing, terminology-constrained generation, and back-translation. Linguistic analyses examined terminology, MWEs, and lexical diversity, including an approach that integrates MWEs into diversity measurement, yielding a domain profile that informed MT experiment design and evaluation. Neural MT and Large Language Models (LLMs) were then fine-tuned under varied data conditions, focusing on compact architectures such as T5, NLLB-200 and Llama-3.1. Evaluation combined automatic metrics with human assessments using the Multidimensional Quality Metrics (MQM) framework and qualitative feedback from professional translators. Findings show that bioinformatics exhibits complex terminology, high MWE density, and a long-tail distribution with a small set of highly frequent terms. Natural corpora display strong lexical diversity, especially after MWE-merging, whereas synthetic data show more variable profiles depending on generation strategy. Domain adaptation substantially improves translation quality on both automatic metrics and human judgment, with domain-adapted models outperforming their baselines and larger models, particularly for terminology and MWE translation. The thesis addresses the under-representation of bioinformatics in linguistic and MT research by providing new corpora, domain-adapted models, and an evaluation framework. Together, these contributions provide both practical tools and conceptual insights for domain-specific MT in bioinformatics, while offering transferable insights for other specialized scientific domains.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Giraud, Jurgi
Rights
- Language dc:language
- en