Recent advances in training deep neural networks have produced  high performing machine reading models.  However, the current success of deep learning hinges upon having large quantities of labeled data to robustly estimate model parameters. For many languages,  little to no labeled data is available for this. 

 

 A key question is therefore: how can we develop machine reading methods whose performance on new languages is not contingent upon availability of substantial amounts of labeled data?  In this talk I will present some of our ongoing work on learning representations that are invariant with respect to the shift in language, and can be used for cross-lingual machine reading.