Video Recordings

Conditional Random Fields (CRFs) were originally proposed for modelling sequential labelling problems, supplanting Hidden Markov Models as the technique of choice for popular tagging tasks such as named entity recognition and information extraction. Since then the underlying formalism of CRFs, conditional modelling with undirected graphical models, has been extended beyond tagging to successfully model a wide variety of more complex structured learning tasks. In this talk I'll describe a number of generalisations of CRFs for modelling problems in Machine Translation (MT). I'll first discuss the problem of finding translational correspondences in parallel corpora (word alignment), a key initial task in building translation models. I'll then describe a translation model that generalises CRFs to parsing with synchronous context free grammars and introduces latent variables in order to cope with the myriad of ways in which source sentences may be translated. These approaches allow the translation process to be conditioned on a wide variety of features of the input sentence, providing considerably more power than current models that rely entirely on local lexical information.