FORUM ON MACHINE TRANSLATION 
Machine Translation will not Work 
Martin Kay 
Xerox Palo Alto Research Center 
3333 Coyote Hill Road 
Palo Alto, CA 94304 
PANELIST STATEMENT 
Large expenditures on fundamental scientific research are 
usually limited to the hard sciences. It is therefore entirely 
reasonable to suppose that, if large sums of money are spent 
on machine translations, it will be with the clear expectation 
that what is being purchased is principally development and 
engineering, and that the result will contribute substantially 
to the solution of some pressing problem. 
Anyone who accepts large (or small) sums on this under- 
standing is either technically naive or dangerously cynical. It 
may certainly be that 
I. machine translation could provide a valuable 
framework for fundamental research; 
2. texts in highly restricted subsets of natural lan- 
guage could be devised for particular puposes and 
texts in translated automatically; 
3. computers have an important role to fill in 
making translations; 
4. translations of extremely low quality may be ac- 
ceptible on occasions. 
However, 
1. the fundamental research is so far from ap- 
plicability, 
2. the language subsets are so restricted, 
3. the useful computer technologies are so different 
from machine translation, 
4. the quality of the translations that can be 
produced of natural texts by automatic means is 
so low, and 
5. the occasions on which those translations could 
be useful are so rare, 
that the use of the term in these cases can only result in con- 
fusion if not deception. 
A determined attempt was made to bring machine trans- 
lation to the point of usability in the sixties. It has become 
fashionable to deride these as "first generation" systems and 
to refer to what is being done now as belonging to the second 
or third generation. It should surely be possible for those 
who think that the newer systems can succeed where the ear- 
lier ones failed, to point to problems that have been solved 
since the sixties that are so crucial as substantially to change 
our assessment of what can be achieved. We know a good 
deal more about programming techniques and have larger 
machines to work with; we have more elegant theories of 
syntax and what modern linguists are pleased to call seman- 
tics; and there has been some exploratory work on anaphora. 
But, we still have little idea how to translate into a closely 
related language like French or German, English sentences 
containing such words as "he", "she", "it", "not", "and", 
and "of". Furthermore, such work as has been done on these 
problems has been studiously ignored by all those currently 
involved in developing systems. 
Unfortunately, the sums that are being spent on MT in 
Europe and Japan are large enough to make virtually in- 
evitable the production of a second ALPAC report sometime 
in the next few years. This will inevitably have a devastat- 
ing effect on the whole field of computational linguistics, 
everywhere in the world. The report will be the more devas- 
tating for the fact that much of the money has in fact been 
spent frivolously, and much of the work has been incom- 
petent, even by today's limited standards. 
268 
