How to improve NLU results when using QNA only?

Hi, I’m using a simple bot that only uses the QNA functionality, currently the bot has about 50 questions, each question has between 4-11 variations/utterances with one answer.

Is there a way to improve the NLU other than loading extra question variations? We have had some good results, but equally some pretty strange ones too, often seeing the intent recognised as “none 100%” - with a valid answer being between 5-10%, see this screenshot:

I’ve asked the bot for information about “childcare” — I have at least 5 questions, each with multiple variations that clearly contain the word childcare (not visible in screenshot), I’d expect the bot to be able to bring back an answer from this using the NLU, not to hit none 100%.

Another example, I have a question with a string of “I want to make an appointment with the Emotional Health Service”, when asking a slight variation of this, clearly with the words “emotional health service”, it hits none 100% again, but you can see the correct answer is being considered at 10%:

This has happened multiple times and I’m not sure how to improve this, when using exact phrasing it works, but we can’t rely on users being exact.

I’ve tried loading my own language server with 300 en dimensions, the logs are not showing any errors, I’ve seen the bot complete it’s training and there are files in the /intents folder so any advice on how to improve this would be much appreciated!


The first and most important thing to note with QnA is that exact match ends at 10 questions. This means that the module will use exact match for any qna with less than 10 questions. Second thing is that your questions have to be about one thing when you draft a single qna. Sounds frustrating but machines do exactly as they are told. So for instance you would have a higher success rate if you include two seperate qna’s for opening times and for closing times as opposed to combining the two topics. I would have loved to see the qna’s but basically it boils down to that.

Bit of a late reply but just to say thanks for this information.

We are using a bunch of QnA entries that contain under 10 questions each, I was under the impression we needed a minimum of 4 to avoid exact matching, so in future i’ll try to get them up to 10!