Understanding the concepts

How the intent condition is evaluated

• any of listed intents are matching
• all of listed entity conditions are matching

So the condition could look like this

( [entity-1] AND [entity-2] ) AND ( [intent-1] OR [intent-2] )

The closest match wins

If there is more then one matching conditions, the closest one wins. Also in 2.xx version bots the sophisticated conditions should be above the less complex conditions.

Also if the difference is too big (lot of additional redundant entities), there'll be no match.

How the matching score is calculated

Let's have a following intent and some entities matched:

{
    "intent": "travel",
    "score": 0.9,
    "entites": [
        {
            "entity": "destination",
            "value": "Prague",
            "score": 1.0
        }
    ]
}

And let's have theese three interactions with following conditions

1. entities: @destination and @transportType, intent: travel

   This interaction is not matching, because there's @transportTypes entity missing.

2. entities: optional @transportType?, intent: travel

   This interaction is matching, but:

   • there'll be little penalisation (0.001) for missing optional entity @transportType
   • there'll be penalisation (0.05) for redundant recognized entity @destination

3. entities: @destination, intent: travel

   This interaction is fully matching and there'll be no penalisation.

4. intent: travel

   This interaction is matching, but:

   • there'll be penalisation (0.05) for redundant recognized entity @destination

5. just entity: @destination

   This interaction is matching, but:

   • there'll be penalisation (0.15) for redundant recognized intent travel

An entity extends the utterance

An utterance should keep its meaning while replacing the marked entity with it’s values.

The bot should be ready to answer all entity combinations

There should be an interaction for all possible combinations of entities used in intent utterances.

Best practices for building the training data

before creating new intent

• use the NLP tester to ensure there is not already similar intent

when adding utterance examples

• always add at least a pair of examples

  example: I need a break and I want a break

• if a word could be omitted, add another example

  example when adding I need a break add also I need break and need break

• combine short examples with long examples (use your imagination)

  example: need coffee and i need a coffee please

• an intent should have at least 10 utterances