Parallelizing Translation

[Alain Scribot]

Hi,

There are several aspects in translation.

1) translation quality level , let's say vs. speed:

Fast: Word for word with hardly any re-reading or after-cross-check...

Medium: Sentence for sentence: Read a sentence, memorize, reproduce in the other
language. Some, but limited after-cross-checking.

(At this level, one can already see the notion of "granularity" coming in. In
parallel processing, one of the questions is often "what's the best "grain" size
for my algorithm", taking into account such things as CPU clock, latency & coms
bandwidth & trying to diminish grain size to optimize performance).

Good: Text to text: Read the whole text, understand it, reproduce in the other
language as though it were your own, using the original just for background
support and as a draft only. (Famous example to illustrate this: translation of
E.A. Poe's tales of M & I in French by C. Baudelaire).

The feeling here is that grain size is increasing rather than decreasing in order
to gain a synthetic general view rather than get lost in details and not convey
the essential.

I  get the impression that machine translation today is somewhere near a bizarre
"Fast" level, (let alone the usual joke of iterating a given text from one
language to the other back & forth a few times)

2) grammar/comprehension: But that's not all, more often than not the initial text
is of pretty poor grammatical quality and of unclear meaning. It may have to be
improved before translation, i.e. it doesn't  mean much even before translation.

 3)context: a text has meaning in a context ("socio-cultural" is a good
description even for engineering matters & not just inches -but unit unstated- to
metric -unit also unstated-  even in English to English translation: ever tried to
order a "pint" in an Aussie bar?-). I don't think we've invented an artificial
socio-cultural context converter yet (an ASCCC?) , but why not.