Note that we’ll be splitting our coverage of this chapter between today and next Wednesday; for today, make sure to have read up to section 11.2, as well as sections 11.7, 11.8, and 11.9.
Warren Weaver’s “original memorandum from 1945” on machine translation is remarkable, and is worth reading for several reasons:
It is incredibly well-written, both in terms of its organization and its rhetoric
(though do take note of a few minor passages that have not aged particularly well- science in the 1940s)
He outlines the general shape of the rule-based MT methods that would be standard for years, and also clearly lays out their limitations
He makes the link between Shannon’s noisy-channel model and the translation problem (about forty years before the computational and data resources made it practical to actually try)
He also foresees that McCulloch & Pitts’s ideas would be crucial (about sixty years early)
His idea of using adaptive context windows for word-sense disambiguation was spot-on
“Oh, yes, everything’s right on schedule, Fred”: Keynote from the “Twenty Years of Bitext” workshop at EMNLP 2013, by Peter Brown and Robert Mercer, on the history of early statistical machine translation.