I've been working on thinking up utterances for my intents, and am having trouble being sure I've covered all the possible utterances, especially when I type each one out by hand, which makes it hard to keep track of ones I'm missing.
The solution seemed to be to use custom slots trained on synonyms to create something like a CFG for my utterances. (e.g. My utterance would be '{verb} the instructions' where the slot is trained on 'list' and 'state' so I have one utterance instead of two). That way I focus on structure and not synonyms.
Now I am working on taking all my utterances and trying to infer a grammar so I can convert them to the more maintainable (custom-slot based) format, but it seems this is a very hard problem (https://en.wikipedia.org/wiki/Grammar_induction). And it hit me, wouldn't this be something Alexa already does? I don't think I should be building grammars on top of grammars...
Recently Alexa added support for synonyms in slot training values, but those aren't advertised as useful for training utterances.
I have a lot of intents, some with fairly similar utterances. How should one make them easy to maintain?