[22] Graham Neubig - Unsupervised Learning of Lexical Information
The Thesis Review
00:00
The Future of Syntax Based Models
In your thesis, the models were these phrase based models. And I think the reason why is because they fundamentally didn't have the ability to handle kind of the hierarchical syntactic structure of language. Now syntax based models did have this ability but the problem is analyzing syntax is hard,. Even defining what you mean by syntax is hard. But then neural MT came out and I would have never believed a priori that neural MT would actually be able to handle syntax properly and generalize things like this. It was a very interesting transition to see the whole field completely turned on its head.
Transcript
Play full episode
