Building natural language parsers

Опубликовано 6 сентября 2016, 16:54

This talk takes a long view on how we approach the problem of parsing natural language. I'll explain how and why what counts as parsing has changed considerably over the past decade or so, what opportunities and applications there are for parsing in the future and what directions research in parsing may take. We'll see how the shift from generative to discriminative models, which was initially motivated in part by the desire to probabilistically model linguistically more realistic grammars in fact lead us to use linguistically less interesting but empirically more accurate models based on MaxEnt and computed in a coarse-to-fine manner. The talk ends with a discussion of current research directions, including self-training and adjoining of hidden states to the parse representation.