Eduard Hovy wrote:
It seems to me the method of construction is less important than the
method of operation, so I'd vote for calling them similar.
This reminds me of Harry Somer's review article on EBMT, where he
suggests that EBMT systems typically exhibit the two following
distinguishing characteristics:
1. Linguistic knowledge is mostly contained in an "example-base", i.e. a
collection of previously translated text, possibly preprocessed to
various degrees;
2. the translation process per se normally proceeds in three distinct
phases, which he refers to as "matching", "alignment/adaptation", and
"recombination".
However, in a follow-up to this article, Davide Turcato and Fred
Popowich argue that by putting too much emphasis on processing issues,
we are missing the point, and we end up with banalities about MT in
general: in the end, "matching" is just another word for "analysis",
"alignment/adaptation" essentially boils down to "transfer", and
"recombination" is mostly a synonym for "generation". (In fact, Somers
himself makes these equivalences explicit, by pasting the EBMT
buzzwords unto the "Vauquois pyramid".) To some extent, *any* MT system
can be claimed to perform all of the above steps. Furthermore, and by
Somers's own account, to carry out each of these tasks, EBMT systems
basically rely on the same plethora of mechanisms as other systems
(taggers, chunkers, parsers, etc.). So basing a definition on this
division of the translation task turns out to be a bit sterile.
Which leaves us, at least in my reading, with only one distinguishing
feature: What truly distinguishes EBMT from other paradigms is its
reliance on examples as a means of capturing and storing knowledge about
language and translation.
Interestingly, by this criterion, word-based SMT (a la IBM) is not EBMT.
But phrase-based SMT is.
Michel
_______________________________________________
Mt-list mailing list