The rise of neural methods and the growth of training data has certainly improved machine translation quality over the last decade. But it is still far from perfect when used "in the wild", e.g., on low-resource languages, very specific domains or non-standard inputs. I want to approach this problem from two perspectives: As a user, how can you teach such a translation system to produce better outputs? And on the machine learning side, how can the translation system efficiently learn from human feedback? The key challenges for such a human-in-the-loop machine learning problem are to find 1) suitable human-machine interaction paradigms, and 2) methods for sample-efficient machine learning. In this talk I will present reinforcement learning algorithms for machine translation that learn from human feedback of various types, their application in real-life, and I will discuss how far these approaches have solved the above challenges.
Julia Kreutzer is a PhD candidate and research assistant in the Statistical NLP Group advised by Prof. Stefan Riezler at the Department of Computational Linguistics at Heidelberg University, Germany. Her research focuses on the integration of reinforcement learning methods into sequence-to-sequence models in order to adapt these models to human preferences.