Eliminate the effect of question mark in utterance
The effect of question mark in utterance prediction should be eliminated, because not all customers write proper questions with question mark, commas, and so on..
example: I write the exact question with (?) gives different intent prediction than the same question without (?)
I don't understand why (?) has this big impact on the prediction
Punctuation no longer has any effect in training or predicting.
Nikki Bolleboom commented
We still experience issues with the effect of question marks in utterances. The predictions are way off when our users write their questions with proper interpunction. Our model uses the Dutch language.
If I may, I humbly recommend that you go research what the properties of a non-deterministic algorithm are, and I think your question will be answered.
Shradha Singh commented
Hi @LUIS UserVoice (Admin)
Would you please help us in understanding
"Of course, LUIS is nondeterministic by nature so it's best not to expect the exact same confidence scores even when the exact same utterance is used twice"
I am going to make use of MS LUIS for a conversational bot. My idea is that, for same query , LUIS will predict same intent every time.Please help in understanding if this is not the case. Because if there is variation in prediction then conversational flow will be disturbed.
A little background:
I am already working on a FAQ bot which has been developed using MS Bot framework and MS QnA Maker. Our team wants to replace MS QnA Maker with MS LUIS.
Hmm, I would think an AI system that accounts for all input (or as much input as possible) is more powerful than an AI system that doesn't. Remember that LUIS operates using confidence scores, so it already is able to recognize questions without question marks. It's just that when a question mark is present, LUIS is more confident about the intent than if a question mark is not present, and that makes sense.
Consider ambiguous input. If a user says "And I'm supposed to click the button?" that's an entirely different intent from "And I'm supposed to click the button." One is asking for clarification and the other is a statement of understanding.
However, it is possible to achieve the functionality you're looking for on the client application's side. You could strip question marks off the end of utterances before they're sent to LUIS and that would achieve the effect of making sure an utterance returns the same intent regardless of whether there's a question mark. (Of course, LUIS is nondeterministic by nature so it's best not to expect the exact same confidence scores even when the exact same utterance is used twice.)