Hi all,
our next speaker is Yacine Jernite from NYU.
The talk will be at AA6214 on August 4, 13:45.
Title: Learning Representations of Language from Text
Abstract:
Learning good general-purpose representations of written language can be useful for a wide range of applications, but there are still a number of open questions in the field, from model architecture to learning strategies. In this talk, we start by briefly going over these questions, then review some recent work related to two of them: the choice of the reading level for text data and of a learning objective for unsupervised training. First, we describe two recent architectures which look at sub-word information to build word and text representations. Then, considering the prevalence of language modeling as an unsupervised objective, we propose an adaptive hierarchical approach which speeds up learning while avoiding some of the pitfalls of comparable methods. Finally, we propose an alternative discriminative learning objective for sentence representation systems based on discourse modeling.
Bio:
Yacine is currently pursuing a PhD at NYU in David Sontag's Clinical Machine Learning group, where his research interests include graphical models, applying machine learning to medical data, and language understanding. His current focus is on using unsupervised methods to take advantage of the vast amounts of freely available text data when learning language representation systems. Yacine holds a Diplome d'ingenieur from Ecole Polytechnique (2011) and an M.Sc. in Machine Learning from ENS Cachan (2012).
Best,