<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
<div class="moz-cite-prefix">On 10/22/2021 2:30 PM, Doug Ewell via
Unicode wrote:<br>
</div>
<blockquote type="cite"
cite="mid:000001d7c78c$168e1c80$43aa5580$@ewellic.org">
<pre class="moz-quote-pre" wrap="">Miscommunication can happen in almost any translation situation, even between two educated, literate, fluent humans, and for that matter even within a single language.
Translating between Spanish and English is supposed to be one of the easiest scenarios in the field, but there is still the Spanish verb 'deber' which can mean either "must" or "should" in English. Getting this right can be tricky; getting it wrong can cause any number of problems.
Achieving 100% perfection is probably never going to happen. Getting within epsilon — using actual software solutions, to translate arbitrary content — happens every day, and the value of epsilon is shrinking.
--
Doug Ewell, CC, ALB | Lakewood, CO, US | ewellic.org
</pre>
</blockquote>
<p>Your argument that humans can't get translations correct or make
them unambiguous should lead you to the conclusion that this
lowers the bar for AI translation.</p>
<p>But that's OK. We may stipulate that the highest achievable
quality can only approach 100% but must fall short.</p>
<p>Even more so, if you go beyond factual accuracy to mood, tone and
style.</p>
<p>Human translators will get the facts wrong, sometimes. I've seen
examples where the switched "left" and "right". A very human
mistake, but then, current AIs will get "he" and "she" wrong at
times.</p>
<p>In proofreading translations, a helpful technique can be to
visualize a description and then compare that with the
visualization of the translated description. You can spot
translators that do not use that technique in creating the
translation quite easily, even if the translation is correct and
well done on the level of words and sentences.</p>
<p>It's definitely something you need in translating fiction, but I
know translators who do this even in fields like banking.<br>
<br>
How do you translate a speech that purposefully includes as many
"r" sounds as possible? Something a famous novelist wrote for a
candidate whose pronunciation of that sound made his speech
particularly attractive.</p>
<p>Well, the number of such texts are really small, so perhaps they
can forever rest in the not fully translatable remainder.</p>
<p>But any ordinary novel usually contains at least one expression
or concept that will really stump a translator, although they may
be rather unremarkable in the original. For example, the
"half-potato" ordered from a street vendor. (Half a link of
sausages with mashed potatos - a shorthand used only regionally
and used to give the story a local flavor).</p>
<p>It's not a term that you'll find in any dictionary, or even in
many online texts. <br>
</p>
<p>Would an AI recognize that it's not use literally? Would it be
able to locate the one travelog where somebody describes it? Would
it be able to condense that information into something usable that
lets a non-local reader get what dish was meant, but not disrupt
the narrative with a clumsy circumlocution?</p>
<p>These kinds of things will be the stumbling blocks in attempting
to approach "100%". Some of them may be amenable to a
redefinition: if your AI promises a "raw translation", it would be
OK to footnote such issues for post-translation. But that is
tantamount to the admission that there's a considerable remnant
that can't be translated without some understanding of meaning and
some theory of mind for the reader of the text (and their
different levels of understanding and familiarity with certain
items, concepts, practices, etc. based on the language and its
associated culture).<br>
<br>
Which was James' premise.</p>
<p>A./<br>
</p>
</body>
</html>