Analyzing and improving cross-lingual knowledge transfer for machine translation

Multilingual machine translation systems learn patterns from parallel text corpora to make knowledge accessible across the world's languages. Understanding how knowledge flows between languages requires navigating subtle nuances in model representations, data availability, and architecture. Harnessing these nuances is challenging because low-resource languages lack the wealth of parallel data that makes transfer comparatively straightforward for their high-resource counterparts.
In this thesis, we analyze and enhance cross-lingual knowledge transfer to deliver robust multilingual NLP capabilities, with machine translation as a central testbed. First, we introduce Representational Transfer Potential to quantify cross-language similarity, uncover predictors like multi-parallel overlap and genetic distance, and design an auxiliary similarity loss that strengthens low- and mid-resource translations. Second, we extend semi-parametric translation with cross-lingual and multilingual k-nearest-neighbor datastores, showing that linguistically structured retrieval boosts quality and enables faster inference without sacrificing accuracy. Third, we examine the fine-tuning paradox in large language models, revealing how parallel data can erode formality control, domain adaptation, and document-level coherence, and we mitigate these losses by blending monolingual and parallel supervision. Fourth, we study the role of language diversity during fine-tuning, demonstrating that scaling the number of translation directions improves both seen and unseen pairs, reduces off-target generations, and aligns representations until gains plateau.
Together, these findings show that deliberate modeling and data choices can extend high-quality multilingual NLP beyond well-resourced languages. Deepening our grasp of cross-lingual transfer not only illuminates current vulnerabilities but also points the way toward inclusive, resilient multilingual technologies that serve the full spectrum of the world's languages.