1898757 - Firefox Translate seems miss-lead by datetime timestamps.

Status: RESOLVED FIXED

(Firefox :: Translations, defect)

Description

[Nicolas B. Pierron [:nbp | PTO until 11/11]]

The following article has a publication timestamps as well as an Updated by: timestamps which are both moved from 2024 to 2019.

The original text contains:

21.05.2024
Обновлено: 22.05.2024

Which is translated (ru -> en) to:

05/21/201924
Updated by: 10:22 May 2019

The corresponding HTML is:

<div>
  <time class="article__datetime timeago" datetime="2024-05-21T19:22:00+03:00" y-use="Timeago">21.05.2024</time>
</div>
<div>
  <span>Обновлено: </span>
  <time class="article__datetime timeago" datetime="2024-05-22T10:08:15+03:00" y-use="Timeago">22.05.2024</time>
</div>

Comment 1

[Paul Adenot (:padenot)]

Interestingly, this doesn't reproduce in about:translations for me when just pasting the second line, but does reproduce partially when pasting the two lines, in the sense that the first line is like in the web page, but the second line is correct.

Comment 2

[Erik Nordin [:nordzilla]]

Attachment: image.png

Thanks for reporting this, Nicolas.

I haven't looked too closely, but this does seem, on the surface, an issue both with the Bergamot engine and with the ru -> en model.

The unfortunate situation is that the Bergamot engine has its own logic for handling translations within markup. We do have an imperfect heuristic for subdividing the DOM into chunks to feed the translator, but it is a weird sort of hybrid situation at present partly due to legacy reasons and the complexity of decoupling it entirely.

In the future, we would love to have all of the DOM processing on our side, and to feed the translator only plain text to translate.

There are definitely improvements to make here, as a whole, and I appreciate this bug report.

---

Paul, you can set the browser.translations.useHTML pref to true and then pass markup directly into about:translations, which does reproduce for me (see attached screenshot).

Comment 3

[Greg Tatum [:gregtatum] (REO Fx132)]

I tested this with our new models, and it seems that the robustness has fixed this issue.

See: https://github.com/mozilla/firefox-translations-training/issues/238

Comment 4

[Greg Tatum [:gregtatum] (REO Fx132)]

We'll still need to retrain older models.