Fidelity and Nuance: Knowledge Integration for Domain-Specific Machine Translation

Sammanfattning

In high-stakes domains such as defence and foreign policy, machine translation (MT) systems must accurately translate specialized terminology while maintaining consistency across long documents. However, both Neural Machine Translation (NMT) systems and Large Language Models (LLMs) often struggle with domain-specific language and context preservation. This thesis investigates how domain knowledge integration and document chunking affect Swedish–English MT in the defence and foreign policy domain. Using glossaries and translation memories (TMs) as external knowledge sources, multiple LLM-based and NMT-based systems were evaluated under increasing levels of knowledge integration. In addition, fixed-size, sentence-aware, and a proposed hierarchical chunking strategy were compared for translating long-form documents and easily integrating relevant external knowledge for specific chunks. The results show that external knowledge generally improves translation quality, with TMs providing larger benefits than glossaries alone. Combined knowledge integration produced the strongest overall performance, while hierarchical chunking consistently improved semantic quality by preserving document-level context. Among the translation systems, larger LLMs achieved the best overall results. These findings show that retrieval-based external knowledge integration and context-preserving chunking can significantly improve domain-specific machine translation without fine-tuning or retraining. This offers a practical approach for translating long defence and foreign policy documents that are rich in terminology.