Storage and computation

David_Tuggy at SIL.ORG David_Tuggy at SIL.ORG
Wed Oct 14 00:05:00 UTC 1998


     Deborah Ruuskanen wrote:

     "Machine translation has tried to use ever[y] larger memories to store
     and retrieve translations once made and match them against
     translations to be made: this method simply does not work if the
     translation to be done is not an *exact* match. So much for retrieval.
     However, if it *is* an exact match, then retrieval saves mucho time
     ..."

     So much for machine retrieval, perhaps. But do we need to posit that
     humans doing retrieval are as literal-minded as computers are? It
     would seem self-evident that we excel at non-exact matching.

     Maybe another way to say it is that we apparently prefer to add a
     little computation to our retrieval system to make it much more
     flexible and efficient, rather than to invest a great deal more
     computation starting over from scratch.

     Once again, if we set up computation and retrieval as either/or
     alternatives, we're setting ourselves on the wrong track. People do
     both, and typically at the same time. Certainly things are weighted,
     as Bill Croft and Wally Chafe and others have been saying, much more
     heavily towards retrieval than the "generative" metaphor would lead
     you to expect.

     A closely related issue: what are we matching, anyway? Probably
     something vastly different from the patterns of 0's and 1's or
     higher-level letters that are all the computer knows. What is not an
     exact phonological (much less phonetic) or even lexico-syntactic match
     may be much more nearly an exact match of the somewhat sloppy semantic
     stuff we're usually primarily comparing in translation. Even the
     phonological and lexico-semantic stuff is almost certainly not be
     stored in as rigid a form as a computer would do it.

     --David Tuggy


______________________________ Reply Separator _________________________________
Subject: Storage and computation
Author:  druuskan at CC.HELSINKI.FI at internet
Date:    10/13/98 10:07 AM



More information about the Funknet mailing list