CODEWITHSUNDEEP
nlpazure-cognitive-servicesazure-language-understanding
JPThorne
JPThorne

Reputation: 610

MS LUIS.ai | Model performance issues | How to increase prediction accuracy

One of our LUIS apps has the following intents, amongst others:

AskAboutFunds & UserGreeting

AskAboutFunds has the utterances funds and fundz (amongst others) mapped to it. Requests for funds come back with 1.0 certainty, fundz however is being incorrectly predicted as belonging to the UserGreeting intent. It's certainty is around 0.68 for AskAboutFunds and 0.79 for UserGreeting, so the higher certainty is returned.

UserGreeting has the utterances "hello", "hi" etc. Those intents are correctly predicted.

So my question is: How do we correct LUIS' errors in this case? I've tried retraining, tried throwing more variations on fundz at it - fund z fundzz funfz etc., but it just never gets more accurate at predicting it even when it's an exact match.

Worth noting - we have Bing spellcheck enabled as well, however, it doesn't correct fundz to funds, unless it's part of a larger utterance, like "what are the available fundz?"

Upvotes: 0

Views: 344

Answers (1)

Ezequiel Jadib
Ezequiel Jadib

Reputation: 14787

If you expect some common misspellings that you expect to be repeated, Luis has something called Phrase List Features that will allow you to define exchangeable (are not exchangeable) words and ultimately to improve the performance of your model.

In this case, I imagine "fundz" being an exchangeable word of "funds".

Here you will find the documentation around Phrase List Features.

enter image description here

Upvotes: 2

Related Questions