Scaling semantic parsers to very large domains

Semantic parsers map natural language sentences to formal representations of their meaning. Recently, algorithms have been developed to learn such parsers for many applications, including question answering, information extraction, and robot control. However, each application requires significant innovations in modeling and learning and we lack semantic parsers that generalize across domains. In this talk, I present a new approach to learning domain independent semantic parsers from indirect supervision, with application to question answering for a wide variety of topics ranging from astronomy to celebrity trivia. The key challenge is to develop new parsing and learning techniques that require very little in-domain data. I will describe a novel two-stage parser that separates a domain independent, linguistically motivated, parse step from a domain dependent ontology matching process. This allows a single grammar to be learnt for multiple different domains while also supporting reasoning about meaning in very large ontologies. Experiments show that the two stage parsing approach significantly outperforms competing systems when answering questions with Freebase. I will also give a taste of ongoing work towards even wider scale natural language interfaces and information extraction systems.

I'd also like to mention the semantic parsing workshop we're putting on at ACL. Website here: http://sp14.ws. We're soliciting short papers with an emphasis on ambition and no need for results. There will be cash prizes for the best of these courtesy of Google. I encourage you and your students to stick something in.