Language technology has transformed the way we write, the way we interact with our devices, and the way we share and consume information. This was made possible by advancements in the field of Natural Language Processing (NLP), a largely data-driven subfield of machine learning. Since data are limited for many of the tasks, domains and languages studied in NLP, transfer learning has gained
great prominence in the field as a way to alleviate data scarcity. This thesis presents work on methods, evaluations and resources for multilingual transfer learning. Our research shows how to improve and correctly evaluate cross-lingual embeddings obtained through alignment. It sheds light on the source of performance in cross-lingual transfer learning for dependency parsing. And it introduces two new resources for language generation tasks, one best viewed as a test bed for cross-domain transfer methods and the other, as a test bed for meta-learning techniques. This thesis contributes to efforts in NLP towards optimal transfer of knowledge across languages and highlights some remaining limitations.