Lecture: 'Wiki, please explain!' by Dr Milica Gasic

Thu 17 May 2018

Trinity Hall

Unlike most of us, computers are awfully good at quickly absorbing huge amounts of information. Unfortunately, they are less good at sharing their knowledge with us. We humans learn best through interactive conversation. And conversations are not something that comes naturally to them.

So, how we do teach computers to talk? How do we get them to tell us what they know? Perhaps, we should hire an ingenious programmer to think of all conceivable questions, come up with clever answers, and type up long lists of these? If we lower our aspirations just a little bit and aim to cater only for, say, "every conceivable question about Cambridge bus times" then such an approach is indeed possible. But your gut feeling probably (correctly) tells you that, in the long run, this is not very promising. There's no way such a rigid question-answer scheme is ever going to lead to an interactive conversation.

In this talk, Milica Gasic explains how researchers in machine learning approach this challenge today. There's been fantastic progress in recent years, as some of you may have noticed. But there's also still a long way to go...

Cost: No charge