HMTA with DeepL mixing formal and informal tone in Spanish

[Dwimble]

I'm using HMTA 10.0 and currently translating a help file from English to Spanish. I'm having a problem with it mixing formal and informal tones and forms of words constantly, within lists of steps and even within the same paragraph. The weird thing is if I go to deepl.com and paste a paragraph there, it doesn't mix the tone...it only happens within HMTA. An AI search said it is a known issue and suggested using the settings or passing custom api parameters if possible to force a formal tone. However, it looks like there are no settings within HMTA to control this, and I can't see any way to pass custom parameters to the API to force a consistent tone. It is doing this so much that it makes HMTA pretty much unusable for Spanish. Is there anything I can do about this?

[Alexander Halser]

Hi Dwimble,

There's not much we can currently do about this. Help+Manual uses the DeepL API version 2. A couple of months ago, v3 was still in beta mode, but they seem to have activated the new version in the meantime. V3 has additional settings about the "tone", v2 had not.

[Dwimble]

Thanks for the info. I know that ChatGPT does support locking in the usted/ustedes formal tone for Spanish via prompt. My company does have an api key for ChatGPT as well, so if I switch to using ChatGPT in HMTA, does it support that in some way? Is there some way to include a prompt to do that? If not, I fear that I'm going to encounter this same problem in French, German, etc with no workaround.

[Tim Green]

It's your regular ChatGPT, but you need to include everything you need in your prompt. Then it should work. This is necessary because it doesn't have your context window in the plugin, so it doesn't "remember" anything from one prompt to the next at the moment.

[Dwimble]

Woah. I hadn't used ChatGPT in it, so I didn't realize it doesn't work like the DeepL plug-in. Am I correct here in seeing that it is just a regular ChatGPT window? So there is no way to preserve any formatting at all? I experimented with it and don't see how I could use that to translate my help files in HMTA. I'd have to manually reformat everything I paste in the window to translate, and then recreate every link, every comment, every table, and so on. This is disappointing, because it looks like there is no feasible way to work around this limitation.

[Tim Green]

Hi Michael,

Have a look here:

https://www.helpandmanual.com/help/hm_w ... dtrip.html

Basically, you need to attach your selection as XML to preserve the formatting when using the ChatGPT plugin.

[Alexander Halser]

So there is no way to preserve any formatting at all?

It does preserve formatting and XML structure including meta information. Please attach the topic selection as XML when prompting ChatGPT and it will return properly formatted XML that you can copy back into the topic (right mouse-click in the chat window).

ChatGPT handles and translates the H&M XML code surprisingly well and usually preserves everything. In those cases where the model returns a generic answer, the text is usually Markdown formatted and H&M handles it as formatted text when you copy the answer back to the editor window. Markdown texts are treated with basic formatting like headings, normal text, lists, etc. When it returns XML code, this is fully formatted containing your original styles and meta information.

What does unfortunately not work is the automatism for bulk conversions like DeepL does. This is the big advantage of the clearly structured DeepL API that ChatGPT does yet not have.

Untitled.png

[Dwimble]

Tim and Alexander,

Thanks for the info! After almost 30 years of being a technical writer, you'd think I'd have first checked the help file after seeing that the ChatGPT window was different than the DeepL window. Alas, apparently even I don't read the freaking manual.

Thanks again! That was helpful and a properly constructed prompt solved the problem.

[Alexander Halser]

a properly constructed prompt

And don't forget: you can store prompts that you might use frequently in the plugin configuration. You don't have to type them every time. The user defined prompts may auto-include the current topic selection as XML.

To use a user defined prompt, right click into the prompt edit window and select from the list.

[Dwimble]

Thanks!

After switching to using the ChatGPT plug-in instead of DeepL, I have found many tedious problems fixed and apparently more reliable translations in some cases. The DeepL plug-in would often get confused by text variables, links, invisible comments, and do other things like mix up formal and informal tones and word forms, randomly delete links after copying the translation into the window, and many other things. It was frequently frustrating and caused errors that were sometimes missed. Switching to the ChatGPT plug-in required some careful prompt construction, but after doing that, it seems to have solved every problem I’ve struggled with in the past and only introduced one incredibly minor formatting problem, which was easily worked around.

It is like having a brand new, excellent product after a year of being regularly frustrated with it. I cancelled my DeepL pro subscription today and will only be using the ChatGPT plug-in from now on. Consequently, I also updated H+M to 10.2 and updated my HMTA. Thanks for the great work!

FYI, in case anyone is interested, here's the prompt I ended up with, which works perfectly for me...

"I am updating the Help file and user manual for an email server and client software. Translate this xml into Spanish and keep the meta information, structure, and links. Use the formal tone (e.g., "usted") for words and verbs when necessary. Do not translate text inside "<draft-comment ...>" tags. Do not translate the text between "<%" and "%>" because those are variables for option and page names that will be automatically translated later."

[Tim Green]

Hi Michael,

Thank you very much for this detailed report. That's really helpful.

[BertholdL]

Thank for sharing your experience. And shame on DeepL! I always thought it to be a reliable tool.

[Tim Green]

Thank for sharing your experience. And shame on DeepL! I always thought it to be a reliable tool.

It is a reliable tool. It has just been temporarily overtaken by other LLMs. This is currently a horse race with extremely rapid changes and development. DeepL was always AI based and you can expect them to catch up or even overtake the others quite quickly. Also, you need to differentiate between translation quality and physical implementation, where DeepL needs to learn about Help+Manual XML. This is not something they support directly, but as agentic AI interfaces develop, you can expect that DeepL will also improve in this area as well.