amazon_jargon Flashcards
what is PMR?
pattern match rule the second step after the exact match NLU
FSTs
Finite state transducers, the step after PMR and exact match in the model.
SLU?
Spoken language understanding.
What are slots vs labels?
The entire NLU pipeline (rules fsts stat model etc.) is meant to take a phrase (utterance) (e.g. dogs are amazing) and output: 1. the domain 2. the intent 3. a label for each word
This might look something like:
INPUT:
play some amazing just in beiber songs from my favorite cool cats playlist from march tenth
OUTPUT:
Music PlayMusicIntent play|Other some|Quantifier amazing|SortType just|ArtistName in|ArtistName beiber|ArtistName songs|MediaType from|Other my|ActiveUserTrigger favorite|PlayListName cool|PlayListName cats|PlayListName playlist|MediaType from|Other march|DateMonth tenth|DateDay
This is definitely a step in the right direction but this is a little too messy for us to just pass to the speechlet. “just in” should probably be “justin” and “march tenth” could be something like “3/10/2017”
It’s also possible that our labels might be too comprehensive and detailed and the speechlet only needs more broad labels, or maybe they can’t digest certain labels at all. It’s clear we need some way to take our label output and convert it to a cleaner set of information to pass to the speechlet.
from https://wiki.labcollab.net/confluence/display/Doppler/Domain+definition+changes+and+testing
What are domain definitions?
What else are domain definitions used for
Domain definitions do more than just determining how we convert our output into something the speechlet can handle. They allow us to keep track of exactly what labels are under which intents and domains. They also allow us to filter out the data that we pull from live data in data mart to train the model also.
from https://wiki.labcollab.net/confluence/display/Doppler/Domain+definition+changes+and+testing