How to train LUIS in right way

Question

I have introduced simple intent - SearchForPerson. So, there are two intents in application 'None' and 'SearchForPerson'. To identify person i am lookin for, I have defined hierarchy entity 'Person' with children 'FirstName' and 'LastName'. Now i am looking for a way how to train LUIS. There are possible approaches:

1. I can upload predefined set of utterances but with different first/last names (100 variation per utterence - "search for {fistname} {lastname}" - for example).
2. I can upload predefined set of utterances and define Phrase list with fistnames and lastnames.
3. Something else.

Can please any one recommend the most efficient way to train LUIS for this simple task. All data will be upload to application by api.

Answer 1

You said the most efficient way in your question:

1. I can upload predefined set of utterances (7) but with different names (100 variation per utterence - "search for {fistname} {lastname}" - for example)

But there are points to rexamine. You shouldn't be attempting to train it with 100 variations, because you're using LUIS so that you don't need to do this. I recommend starting off with at most a dozen or so utterances.

The number of utterances to train it with will vary, especially depending on which language you're using. I'm assuming that you're using the en-us culture. Because of this, it will more easily recognize common names in English than in, say Danish or Telugu.

To import your dataset, you'll go to Train & Test in your LUIS app. Here's a screenshot of what the page will look like where you can import your utterances:

Lastly, here is the accepted dataset syntax information from the LUIS portal:

Dataset allowed syntax

The file should be of JSON format.

The file should contain an array of utterances that follow the following JSON format:

[
  {
    "text": "hey dad, are you hungry?",
    "intent": "None",
    "entities": 
      [
        {
          "entity": "FamilyMember",
          "startPos": 4,
          "endPos": 6
        }
      ]
  },
  {
    .
    .
    .
  }
]