Translating pronouns is a challenge that current machinetranslation systems are ill-prepared to handle because it requires complexcross-sentence dependencies in the target language and because the pronounsoften can't be translated literally even in a very literal translation style.My talk is going to focus on the problem of modelling pronominal anaphora inthe framework of phrase-based SMT. Pronoun translation is first cast as aclassification task that is separate from SMT. I describe a neural networkclassifier for this task that models anaphoric links as latent variables andcan be trained on parallel bitexts without explicit coreference annotations.Then, I present Docent, my document-level local search decoder for phrase-basedSMT, and show some results from the integration of the anaphora classifier inthis framework.