The paper argues that the IBM statistical approach to machine translat
ion has done rather better after a few years than many sceptics believ
ed it could. However, it is neither as novel as its proponents suggest
nor is it making claims as clear and simple as they would have us bel
ieve. The performance of the purely statistical system (and we discuss
what that phrase could mean) has not equalled the performance of SYST
RAN. More importantly, the system is now being shifted to a hybrid tha
t incorporates much of the linguistic information that it was initiall
y claimed by IBM would not be needed for MT. Hence, one might infer th
at its own proponents do not believe 'pure' statistics sufficient for
MT of a usable quality. In addition to real limits on the statistical
method, there are also strong economic limits imposed by their methodo
logy of data gathering. However, the paper concludes that the IBM grou
p have done the field a great service in pushing these methods far fur
ther than before, and by reminding everyone of the virtues of empirici
sm in the field and the need for large scale gathering of data.