Unknown said:

Thanks for your efforts, but this is generic nonsense out of a can.

It's only a computer!  Why not try seeing if you can provide more specific details to get a more sensible answer.  Knowing exactly why you have this problem is tricky even for a human skilled enough to know (not me) where the issue might be.  But ultimately the TradosAI will only answer based on what you tell it.  It's not smart enough to read between the lines, ignore misinformation, fill in the gaps etc.  There are also many good reasons for this which I intend to elaborate on soon enough when I get some time.

In the meantime...

- are all your XML files the same structure?

- do you have all the images?

- checked for typos?

- have you validated the stylesheet in another tool?

- can you share the files perhaps?

Paul 

Paul said:

It's only a computer!

I know. I was trying it out the first time. I think it could be useful for frequently asked questions. If there is a way to analyse the content of the original post and assess how likely it is that this AI system can generate a helpful answer, that would be good. You could then suggest using the AI to the user and at other times de-emphasize it. Also, might be good to tag it as "suggested answer", not as "verified answer".

Concerning the XSLT:

a) Yes, all XML files have the same structure.

b) The images are retrieved from the web

c) Not clear how a typo could affect this - images that don't load initially do load after restarting Trados

d) I think I ran it through Saxon

e) Yes, but not publicly. The error only occurs after a few dozen files, never right away.

Daniel

Unknown said:

Also, might be good to tag it as "suggested answer", not as "verified answer".

Indeed... but the idea is that you engage it.  If it doesn't answer your question then reject it and use the reply to refine the answer and have another go as you can see from this page:

 (Missing Wiki Page) 

We're a bit restricted with the platform on how we can implement this, and training it it is very difficult for many reasons:

• some users are not good at asking questions
• some users often mark answers as verified anyway even though they are not even answering the question
• some users mark inaccurate answers as verified
• some users never tell us anything after getting a reply (another reason why we mark it as verified)
• product help documentation rarely helps with the sort of questions we see in here so we don't have a lot to train with
• users (including RWS) often help people off the forum and never come back and explain what they did to help
• The KB is quite good, it's at least structured nicely with Q&A, but it's not enough

So we have a large exercise ongoing to try and improve this, but it's made even harder to see how effective we get because it's hard to get users to try it out and persist in learning how to work with it to help themselves. In the testing I've played with you can improve the question quite often and get better and better answers.  It's as moch about providing the right "prompt" as it is feeding it with the right data to train it more specifically.

But this is early days and we'll persist!

Unknown said:

Not clear how a typo could affect this - images that don't load initially do load after restarting Trados

I have no idea how you compiled your XSLT as you haven't shared it so I don't know what it's doing.  I can only try and guess a bit wider than the AI :-)

Unknown said:

The error only occurs after a few dozen files, never right away.

ok - so you have one project with more than 50-60 files in it and you open them one at a time or all at once?  After 50-60 files it starts getting upset?

Feel free to send them my way if you like?

Unknown said:

Also, might be good to tag it as "suggested answer", not as "verified answer".

Indeed... but the idea is that you engage it.  If it doesn't answer your question then reject it and use the reply to refine the answer and have another go as you can see from this page:

 (Missing Wiki Page) 

We're a bit restricted with the platform on how we can implement this, and training it it is very difficult for many reasons:

• some users are not good at asking questions
• some users often mark answers as verified anyway even though they are not even answering the question
• some users mark inaccurate answers as verified
• some users never tell us anything after getting a reply (another reason why we mark it as verified)
• product help documentation rarely helps with the sort of questions we see in here so we don't have a lot to train with
• users (including RWS) often help people off the forum and never come back and explain what they did to help
• The KB is quite good, it's at least structured nicely with Q&A, but it's not enough

So we have a large exercise ongoing to try and improve this, but it's made even harder to see how effective we get because it's hard to get users to try it out and persist in learning how to work with it to help themselves. In the testing I've played with you can improve the question quite often and get better and better answers.  It's as moch about providing the right "prompt" as it is feeding it with the right data to train it more specifically.

But this is early days and we'll persist!

Unknown said:

Not clear how a typo could affect this - images that don't load initially do load after restarting Trados

I have no idea how you compiled your XSLT as you haven't shared it so I don't know what it's doing.  I can only try and guess a bit wider than the AI :-)

Unknown said:

The error only occurs after a few dozen files, never right away.

ok - so you have one project with more than 50-60 files in it and you open them one at a time or all at once?  After 50-60 files it starts getting upset?

Feel free to send them my way if you like?