23 November 2018 - Jonathan Berant : Seminar
Training from denotations
A common training scenario in natural language understanding, termed training from denotations, is the following: given (input, output) pairs a model must learn to map the input into a discrete latent representation from which the output can be deterministically computed. Training from denotations is useful for mapping questions into SQL queries, question answering with a search engine, mapping instructions to programs, reasoning over knowledge-bases and more. Training from denotations raises two challenges. First, the model must search in a large combinatorial discrete space, sparsely populated with good target states. Second, the model must find paths to target states that generalize well at test time. In this talk, I will describe recent work tackling these challenges. First, I will describe a search algorithm for mapping instructions to programs, where denotations are used at training time to train a critic network from which a good search heuristic can be learned. Second, I will describe a new policy gradient algorithm, where a memory buffer is used to overcome some of the issues of training with policy gradient, such as cold start and high variance gradient estimation, which allows benefiting from the exploration properties of policy gradient. Time permitting, I will also discuss recent work in which we combat spuriousness issues working over an abstract representation and using a cache to share information between examples in the training set.
Jonathan Berant is a senior lecturer at The Blavatnik School of Computer Science in Tel-Aviv University, specializing in Natural Language Understanding. He holds a Ph.D in Computer Science from Tel-Aviv University, was a post-doctoral fellow at Stanford University from 2012-2015, and a post-doctoral fellow at Google Research from 2015-2016. He won best paper awards at ACL and EMNLP, and is a recipient of multiple grants including an ISF, BSF, Samsung runway award, a Joy grant, and an ERC starting grant.