 Hello, my name is Dmitri Bardano and I'm a PhD student at Montreal Institute of Learning Algorithms where my advisor is Yoshio Benjo. In this video, I will present the latest project that I worked on which is called Learning to Compute Board meetings from the fly and which is a joint work with people from Yelonio University Montreal Institute of Learning Algorithms in DeepMind. The problem that we address is the fact that there is a lot of rare phenomenon in language. For example, there is a lot of rare words and the distribution of rare words has a very heavy tail. There can be idioms in the language that they are rare and named and this geographical and these are rare. So prior to this work, the approaches to deal with the sparsity were either to use restricted vocabulary and learn embeddings only for a short list of words or used embeddings for training on vast corpora. The alternative approach that we investigated proposes to use dictionary definitions. We try to add another network to our main network that is trained to read the dictionary definitions and produce an embedding which is useful for the main network to solve the downstream task. For example, here there is a rare word Archegonium and the definition reader is trained to produce a useful embedding of this word. And the whole system is trained together and to end. We evaluated our approach on reading comprehension, which is a task in which a question about the passage should be answered. And we found that using dictionary definitions allows us to narrow the gap between the baseline approach and using embeddings trained on 840 billion words. And we found that the improvement dictionary definitions brings adds up with the improvement that using spelling brings. And we were able to see how having a dictionary definition, for example, for the word Autumn allows the model to match this word with the word Season and without dictionary, a similar model cannot do it. We also evaluated our model on the task of text challenge table, which is and we found that it is also helpful to use definition in this problem. And we found, as expected, that most of improvement comes from examples that include rare words.