Supervisor

Suitable for

Abstract

Discriminative language modelling (hard) Language models are key to many emerging technologies such as speech recognition and machine translation. Most current systems use generative models which are very fast but only capture limited features of the language. Discriminative models have been proposed as a way of introducing much richer conditioning into language models, but all previous approaches have proved too slow for real world use. In this project we'll try and circumvent these performance issues by training discriminative language models using a pseudo-likelihood objective function and stochastic sampling techniques.