<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
<div class="moz-cite-prefix">On 10/23/2021 10:04 PM, Doug Ewell via
Unicode wrote:<br>
</div>
<blockquote type="cite"
cite="mid:003801d7c894$a13132b0$e3939810$@ewellic.org">
<blockquote type="cite" style="font-size: large; color: #3333FF;">
<pre class="moz-quote-pre" wrap="">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.
</pre>
</blockquote>
<pre class="moz-quote-pre" wrap="">As will some native speakers of Mandarin. (他 (he), 她 (she), and 它 (it) are all pronounced "tā" in Mandarin, which blurs the distinction between these three pronouns when they switch to English.)
</pre>
</blockquote>
<p><font face="Candara">Interesting. <br>
</font></p>
<p><font face="Candara">My example was something I encountered for
European languages (although, I didn't tell you that).</font></p>
<p><font face="Candara">Current state of the art seems to limit the
context to a single sentence(*), and one where both source and
target language mark gender in pronouns.</font></p>
<p><font face="Candara">Leads you to suspect that pattern matching
can override the choice of pronoun, which you might think is a
bit weird, but not if you consider languages with both
grammatical and natural gender.</font></p>
<p><font face="Candara">In German, you could write this (translated
verbatim):</font></p>
<p><font face="Candara">We miss our cat. She died yesterday.
<Some male name> will not be forgotten. At least he lived
a long and happy life.</font></p>
<p><font face="Candara">That works because "cat" (Katze) has female
grammatical gender, but the individual animal has natural
gender. When you shift the focus by using the name, you'll need
to switch gender for the pronoun.</font></p>
<p><font face="Candara">There may be even better examples of this.
But if your AI doesn't know about a world outside a sentence,
there will be translation patterns where a male pronoun in the
source becomes a female pronoun in the target somewhat randomly.</font></p>
<p><font face="Candara">If you went to longer contexts you might be
able to account for that, or at least not have a random gender
walk in the translation. But if you do, you are only getting
better on the question of accuracy(**).</font></p>
<p><font face="Candara">You've done nothing for mood, style, and
tone if these aren't directly captured verbatim.</font></p>
<p><font face="Candara">A./</font></p>
<p><font face="Candara">(*) I doubt this even covers all single
sentences - for likely exceptions, see the works of Kant.</font></p>
<p><font face="Candara">(**) someone mentioned that human
translators don't always agree on the best translations. And,
sometimes authors who are not bilingual feel emboldened to
insist on a more "literal" translation, because to their limited
understanding of the target language, a more natural phrasing
may sound unusual and there for different from what they think
it should have been. Therefore, what constitutes an "accurate"
translation is very much a moving target.<br>
</font></p>
</body>
</html>