Responses to Current Research in Machine Translation by Harold Somers

Abstract

"Current Research in Machine Translation" is a sound presentation of issues from the perspective of long experience with the affirmed MT architectures, and the presumed directions of evolution from one architecture to the next. Somers carries this experience, despite its disappointments, into an enthusiastic outlook of a different set of architectures. There are several points, two in particular, made by Somers' paper which I found particularly gratifying. These assertions agree with some conclusions I had come to recently, and held rather closely, if not privately. These have to do with what is or can be non-linguistic about MT. The first of these points is suggested by the accurate observation that in the (now 25) years since ALPAC, MT does not produce radically improved results over the output available then. I think anyone who compares modern output (both from research and from commercial systems) with the ALPAC set, either from a linguistic coverage or correctness perspective, will come to a similar conclusion. Yet, this is not to say that MT is therefore as useless as it was in 1966. On the contrary, MT has an established place in the commercial world, and a relatively long-term commitment from a variety of governments. This is because, in 1966, input was submitted to attendants of a monolithic "computer" and output was provided in the form of (at best) printouts on a green-bar. There was no way that the information consumer of that time could be expected to modify the output to make it legible. The advent of a user environment, and all that has followed from that, has led to word processing tools that make bad output editable into useful output, in ways that the ALPAC writers could not have envisioned. Thus the good news is that MT quality today is usable; the bad news is that the reason that it is usable has much less to do with MT research than we would like to claim. This is I think, consistent with Somers' point. At the same time, though, I think that there is another side to the quality/usefulness discussion that bears some delving into. MT, and natural language in general, is not going to be used for as many things as the ALPAC-era people may have thought. We thought we would use speech as the primary user interface to computers, and, in advance of that, natural language interfaces. Neither of these appear likely, not only because we do not know how, but because the interim, alternate interface modalities turn out to be better solutions.