Why content filtering for chinese not working for OpenAI chat api

Question

Why content filtering for chinese not working for OpenAI chat api

Danny Chooi 0

We need to detect and filter out harmful message from the input and thus we try to utilize the content filtering from azure openai service
When I test with my deployment in the playground for chat, the harmful message could correctly trigger the content filtering, however, when I use an api call with exact same parameters and prompt, the content filtering stop working, and the prompt is deemed as safe, which is different from what we get form the playground.

Could we know why this would happen?

Deleted

This comment has been deleted due to a violation of our Code of Conduct. The comment was manually reported or identified through automated detection before action was taken. Please refer to our Code of Conduct for more information.
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-09-07T15:17:17.2466667+00:00

Danny Chooi Greetings & Welcome to Microsoft Q&A forum!

when I use an api call with exact same parameters and prompt, the content filtering stop working, and the prompt is deemed as safe, which is different from what we get form the playground.

Can you please provide with more information about the API call you're making?

Specifically, can you share the exact parameters and prompt you're using in the API call? This will help better understand the issue.

Also which model are you using?
Danny Chooi 0 Reputation points

2023-09-08T01:05:07.2733333+00:00

Thanks for reaching out, Chinese is the main language we use, and such message "橫又死掂又死，堅持落去都無意思" could trigger the content filtering, which is ideal for our use case. However, when I copy the script directly using the "view code" btn and run the api call locally, such message would not trigger content filtering, instead, the content filtering would refer my prompt as safe. I noticed that when I prompt it to translate such message to english, it would trigger content filtering. Is content filtering language specific?

Thanks for your answering

parameters.txt
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-09-08T06:29:29.37+00:00

@Danny Chooi Thanks for the additional details.
I will check and update you on this.

Appreciate your time and patience.
Danny Chooi 0 Reputation points

2023-09-21T13:18:13.9066667+00:00

hi @AshokPeddakotla do you have any update after research of past week. Highly appreciate if you can get back and help
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-10-05T13:26:08.0666667+00:00

Apologies for the delayed response. As specified in Content filtering,

The content filtering models have been specifically trained and tested on the following languages: English, German, Japanese, Spanish, French, Italian, Portuguese, and Chinese. However, the service can work in many other languages, but the quality may vary. In all cases, you should do your own testing to ensure that it works for your application.

In addition to the content filtering system, the Azure OpenAI Service performs monitoring to detect content and/or behaviors that suggest use of the service in a manner that may violate applicable product terms. For more information about understanding and mitigating risks associated with your application, see the Transparency Note for Azure OpenAI. For more information about how data is processed in connection with content filtering and abuse monitoring, see Data, privacy, and security for Azure OpenAI Service.

It's possible that there is some other difference between the Playground and your local environment that is causing the discrepancy.

To troubleshoot this issue further, I would recommend checking the API version and configuration you are using to ensure that it is consistent with the Playground. You can also try adjusting the configuration of your API call to see if that affects the content filtering behavior. In my observation ,changing the api version to latest might work.

Can you try with different api version and see if that makes any difference?

Let me know if that helps.
YutongTie-MSFT 53,971 Reputation points Moderator

2023-10-10T16:56:04.7833333+00:00

@Danny Chooi Thanks for your feedback again, do you have a chance to check on Ashok's response, please let us know if you need more help. Thanks.

1 answer

Your answer

Deleted

This comment has been deleted due to a violation of our Code of Conduct. The comment was manually reported or identified through automated detection before action was taken. Please refer to our Code of Conduct for more information.
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-09-07T15:17:17.2466667+00:00

Danny Chooi Greetings & Welcome to Microsoft Q&A forum!

when I use an api call with exact same parameters and prompt, the content filtering stop working, and the prompt is deemed as safe, which is different from what we get form the playground.

Can you please provide with more information about the API call you're making?

Specifically, can you share the exact parameters and prompt you're using in the API call? This will help better understand the issue.

Also which model are you using?
Danny Chooi 0 Reputation points

2023-09-08T01:05:07.2733333+00:00

Thanks for reaching out, Chinese is the main language we use, and such message "橫又死掂又死，堅持落去都無意思" could trigger the content filtering, which is ideal for our use case. However, when I copy the script directly using the "view code" btn and run the api call locally, such message would not trigger content filtering, instead, the content filtering would refer my prompt as safe. I noticed that when I prompt it to translate such message to english, it would trigger content filtering. Is content filtering language specific?

Thanks for your answering

parameters.txt
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-09-08T06:29:29.37+00:00

@Danny Chooi Thanks for the additional details.
I will check and update you on this.

Appreciate your time and patience.
Danny Chooi 0 Reputation points

2023-09-21T13:18:13.9066667+00:00

hi @AshokPeddakotla do you have any update after research of past week. Highly appreciate if you can get back and help
AshokPeddakotla-MSFT 35,971 Reputation points Moderator

2023-10-05T13:26:08.0666667+00:00

Apologies for the delayed response. As specified in Content filtering,

The content filtering models have been specifically trained and tested on the following languages: English, German, Japanese, Spanish, French, Italian, Portuguese, and Chinese. However, the service can work in many other languages, but the quality may vary. In all cases, you should do your own testing to ensure that it works for your application.

In addition to the content filtering system, the Azure OpenAI Service performs monitoring to detect content and/or behaviors that suggest use of the service in a manner that may violate applicable product terms. For more information about understanding and mitigating risks associated with your application, see the Transparency Note for Azure OpenAI. For more information about how data is processed in connection with content filtering and abuse monitoring, see Data, privacy, and security for Azure OpenAI Service.

It's possible that there is some other difference between the Playground and your local environment that is causing the discrepancy.

To troubleshoot this issue further, I would recommend checking the API version and configuration you are using to ensure that it is consistent with the Playground. You can also try adjusting the configuration of your API call to see if that affects the content filtering behavior. In my observation ,changing the api version to latest might work.

Can you try with different api version and see if that makes any difference?

Let me know if that helps.
YutongTie-MSFT 53,971 Reputation points Moderator

2023-10-10T16:56:04.7833333+00:00

@Danny Chooi Thanks for your feedback again, do you have a chance to check on Ashok's response, please let us know if you need more help. Thanks.

Answer 1

Teru Lei 10

I have similar problem. I tried to update my model used in my testing Simplified Chinese Q&A bot which worked smoothly some time ago from 3.5 turbo to gpt-4-32k, after that I found the chatbot not work and finally I got that it's because the input is filtered by content filtering. However the content does not include any harmful information at all (The bot is for general information inquiry for hotel). I change the input to English with same meaning then the bot can work again. The default content filtering engineer has problem and please fix it. Luckly the bot is not under production.

唐飞 10 Reputation points

2023-10-16T07:11:55.63+00:00

same here, plz fix it.
慶飛魏 0 Reputation points

2023-11-20T11:08:44.5266667+00:00

same here, plz fix it.

Share via

Why content filtering for chinese not working for OpenAI chat api

1 answer

Your answer