Can Cognate Prediction Be Modelled as a Low-Resource Machine Translation Task?