Title: Deep Learning: Building Models beyond English through Bilingual Learning

Abstract: Deep Learning based methods have shown promise in recent times in solving problems in the area of natural language understanding. The research shows that DNN based methods have comparable performance for many problems and in some cases better results than the existing methods. A key requirement for production quality DNN based methods is availability of large amounts of data. This requirement makes learning models for languages apart from English challenging. In this talk, we would like to briefly introduce various methods that have been studied in literature to handle the problem by employing bilingual training. The talk will begin with brief introduction of language modeling using DNN and then discuss various methods of incorporating bilingual training. We will discuss methods of learning transformations between latent spaces for two languages, learning joint embedding and various other techniques being proposed in literature.


Speaker Bio: Rahul Agrawal is Principal Research Manager at Microsoft Bing Ads, where he is responsible for query understanding and ad understanding for Bing. His primary focus is on driving the charter for non-english languages. He leads a team of 15 scientists and engineers at Microsoft. His primary research interests include deep learning, large scale graph mining, topic models and large scale machine learning. Prior to Microsoft, he was with Yahoo Labs, where he worked on large scale click prediction for display advertising.