Statistical Machine Translation
Miðeind's Erlendur System for Machine Translation Tasks
Pages
6
Time to read
17 mins
Publication
Language
English
Pages
6
Time to read
17 mins
Publication
Language
English
This technical report describes Miðeind's contributions to the Tenth Conference on Machine Translation (WMT25), specifically focusing on the General Machine Translation Task and the Terminology Translation Task. The report outlines the Erlendur system, a multilingual LLM-based translation service designed to enhance translation quality and grammatical accuracy for English to Icelandic translations. Erlendur employs a multi-stage pipeline approach, which includes preparatory text analysis, dictionary lookups, glossary integration, and grammatical error correction. The system's performance is evaluated through its rankings in the WMT25 tasks, where it achieved notable positions, including 3rd-4th in the General Machine Translation Task and 1st in the more demanding Track 2 of the Terminology Translation Task. The report details the system's architecture, model selection, and the methodologies used for glossary integration and context handling, highlighting its capabilities in addressing challenges faced by lower-resource language pairs.