Document-Level Machine Translation

This thesis transforms how computers translate between languages by considering entire documents rather than isolated sentences. Traditional translation systems often miss connections between sentences, creating inconsistencies in pronouns, terminology, and style. This thesis develops innovative solutions including: specialized data enhancement techniques, efficient model architectures that process longer texts, adaptations of large language models for translation, and a collaborative multi-agent system mimicking professional translation workflows. These advances help preserve meaning across languages while maintaining consistency throughout documents, which is critical for literature, technical manuals, and international communication. The work represents a significant step toward translation technology that better captures the interconnected nature of human language.