We have a skill with a few intents and one custom slot type "ITEM" with hundreds of values. One of the intents has a single slot utterance:
ItemItent: "{item}" where item has the type ITEM.
When the user invokes the ItemIntent with a value from the ITEM slot type Alexa sends a request with the slot which contains the value (what the user actually said) and the slot resolution:"item": { "name": "item","value": "tomatoes", "resolutions": { "resolutionsPerAuthority": [ ... ] }
}When the user says a value that is not yet in the ITEM type it also triggers the item intent where the value is set but the resolution is empty (status code ER_SUCCESS_NO_MATCH). This is estimable for us because we want the user to be able to say whatever item they want.
Now we extended the skill with more utterances for the existing intents and with more intents. This changed the speech recognition.
Even though Alexa would still invoke the ItemIntent when the user said something that was not yet in the ITEM type the value and the resolution thing was lost (not even the empty resolution was there). (We tested with the same values before and after the changes to the model.)
At the same time we are sure that Alexa understood the word correctly before and after the changes. We could see it in the history in the Alexa app when we were using Echo devices to test as well as in the Alexa test console's "chat".
As a result we had to decide not to extend the skill with fresh utterances and features in favor of the better "item" recognition.
The question is: Why is the content of the Alexa request changing when Alexa clearly understands what the user said and even maps it to the ItemIntent?
Source: amazon.com