Re: Asking for help: Update German translation of NR using LLM

--0000000000006bb3e306573a23da
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

the state of LLM has changed a lot over the past year since that evaluation
write-up.
I use LLM/NLP every day for my work, and the leaps in improvement is
impressive.
You'll need to be careful for sure. But I suggest  you try it. Do it using
2 separate LLM agents (different models, not just different model versions)
and use a 3rd to evaluate the 2 translations for semantic match and
conflict resolve against original text. (just a suggestion, but you can
play with it until you find a reliable workflow). Once the desired workflow
is established, you can automate the pipeline. If all the documents you
have are related in subject matter, you can create a feedback loop to help
it improve.
In short, if you only have 1 document to translate, then this won't be a
time saver. But if you have a large corpus to translate (especially if it's
a narrow domain) then a workflow like this can be a significant time
savings.
Good luck.
Charlie

On Wed, Jul 22, 2026 at 12:52=E2=80=AFPM Dr. Arne Babenhauserheide <arne_ba=
[email protected]>
wrote:

> Werner LEMBERG <[email protected]> writes:
>
> > I thus wonder whether a guy with LLM experience could help with
> > updating the German translation of the Notation Reference =E2=80=93 or =
rather,
> > generating a complete translation from scratch, since the German
> > translation is severely outdated.  Today, websites like Google
> > Translate provide translations (both German-to-English and
> > English-to-German) that are nearly perfect even for quite complicated
> > grammatical structures, which still amazes me.
>
> My experience is that such a translation looks nearly perfect, but the
> actual content contains hard to spot errors that can even inverse the
> meaning of some sections. I tried that for an article of mine and did a
> full evaluation. The result was that after I fixed all the errors (I
> found), I didn=E2=80=99t actually save time -- despite being the author o=
f the
> original text: https://www.draketo.de/software/ai-translation-evaluated
>
> The worst errors (completely changed meaning):
> https://www.draketo.de/software/ai-translation-evaluated#completely-chang=
ed
>
> Finding those errors is extremely hard, because the LLM is great at
> fudging the text around the errors such that they sound like a correct
> part of the text.
>
> > of technical terms will be mistranslated; we thus have to proofread
> > the results (I volunteer for that).
>
> You may actually be faster (and the result better) if you directly
> translate the new text with the old translation as reference.
>
> Except if the bottleneck for you is typing speed -- if you can read
> English and German side by side much faster than you type one of them,
> then an LLM may help by reducing the amount of text you need to type).
>
> Best wishes,
> Arne
> --
> Unpolitisch sein
> hei=C3=9Ft politisch sein,
> ohne es zu merken.
> https://www.draketo.de
>

--0000000000006bb3e306573a23da
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr">the state of LLM has changed a lot over the past year sinc=
e that evaluation write-up.<div>I use LLM/NLP every day for my work, and th=
e leaps in improvement is impressive.</div><div>You&#39;ll need to be caref=
ul for sure. But I suggest=C2=A0 you try it. Do it using 2 separate LLM age=
nts (different models, not just different model versions) and use a 3rd to =
evaluate the 2 translations for semantic match and conflict resolve against=
 original text. (just a suggestion, but you can play with it until you find=
 a reliable workflow). Once the desired workflow is established, you can au=
tomate the pipeline. If all the documents you have are related in subject m=
atter, you can create a feedback loop to help it improve.</div><div>In shor=
t, if you only=C2=A0have 1 document to translate, then this won&#39;t be a =
time saver. But if you=C2=A0have a large corpus to translate (especially=C2=
=A0if it&#39;s a narrow domain) then a workflow like this can be a signific=
ant time savings.</div><div>Good luck.</div><div>Charlie</div></div><br><di=
v class=3D"gmail_quote gmail_quote_container"><div dir=3D"ltr" class=3D"gma=
il_attr">On Wed, Jul 22, 2026 at 12:52=E2=80=AFPM Dr. Arne Babenhauserheide=
 &lt;<a href=3D"mailto:[email protected]">[email protected]</a>&gt; wrote:<br><=
/div><blockquote class=3D"gmail_quote" style=3D"margin:0px 0px 0px 0.8ex;bo=
rder-left:1px solid rgb(204,204,204);padding-left:1ex">Werner LEMBERG &lt;<=
a href=3D"mailto:[email protected]" target=3D"_blank">[email protected]</a>&gt; writes:<b=
r>
<br>
&gt; I thus wonder whether a guy with LLM experience could help with<br>
&gt; updating the German translation of the Notation Reference =E2=80=93 or=
 rather,<br>
&gt; generating a complete translation from scratch, since the German<br>
&gt; translation is severely outdated.=C2=A0 Today, websites like Google<br=
>
&gt; Translate provide translations (both German-to-English and<br>
&gt; English-to-German) that are nearly perfect even for quite complicated<=
br>
&gt; grammatical structures, which still amazes me.<br>
<br>
My experience is that such a translation looks nearly perfect, but the<br>
actual content contains hard to spot errors that can even inverse the<br>
meaning of some sections. I tried that for an article of mine and did a<br>
full evaluation. The result was that after I fixed all the errors (I<br>
found), I didn=E2=80=99t actually save time -- despite being the author of =
the<br>
original text: <a href=3D"https://www.draketo.de/software/ai-translation-ev=
aluated" rel=3D"noreferrer" target=3D"_blank">https://www.draketo.de/softwa=
re/ai-translation-evaluated</a><br>
<br>
The worst errors (completely changed meaning):<br>
<a href=3D"https://www.draketo.de/software/ai-translation-evaluated#complet=
ely-changed" rel=3D"noreferrer" target=3D"_blank">https://www.draketo.de/so=
ftware/ai-translation-evaluated#completely-changed</a><br>
<br>
Finding those errors is extremely hard, because the LLM is great at<br>
fudging the text around the errors such that they sound like a correct<br>
part of the text.<br>
<br>
&gt; of technical terms will be mistranslated; we thus have to proofread<br=
>
&gt; the results (I volunteer for that).<br>
<br>
You may actually be faster (and the result better) if you directly<br>
translate the new text with the old translation as reference.<br>
<br>
Except if the bottleneck for you is typing speed -- if you can read<br>
English and German side by side much faster than you type one of them,<br>
then an LLM may help by reducing the amount of text you need to type).<br>
<br>
Best wishes,<br>
Arne<br>
-- <br>
Unpolitisch sein<br>
hei=C3=9Ft politisch sein,<br>
ohne es zu merken.<br>
<a href=3D"https://www.draketo.de" rel=3D"noreferrer" target=3D"_blank">htt=
ps://www.draketo.de</a><br>
</blockquote></div>

--0000000000006bb3e306573a23da--