Meeting minutes
dirk's walkthrough updates
dirk: does walkthrough match description?
… this is just the subset that matches part of the text
Action: debbie to update text description in Section 4
dirk: reviews Figure 3
dirk: combination of local stuff and remote
… knowledge graph can be used to find location given GPS
dirk: ASR is sent in parallel to both NLU's (local and remote) (step 6)
jim: what information is given to provider registry?
… how does the provider registry work?
dirk: doesn't know anything about utterance semantics
… it just tries to determine the number of registered IPA providers
… it can use context
… or user preferences
jim: what do input and output from provider registry look like?
dirk: provider selection service gets the ASR result
debbie: what information is in step 9
dirk: it's always the same, just a query for all providers
… the provider registry could be more or less smart
jim: after provider selection service gets an ASR result it queries the Provider Registry for all providers
debbie: what about steps 10 and 11?
dirk: for each provider returned from the registry, authentication information is returned
jim: suggest sending NLU to provider selection service instead of ASR result
dirk: or could just send the raw audio
jim: now the Provider Registry sends all the providers, if we use NLU we could be more selective about what providers we return
dirk: this is the provider selection strategy
dirk: this is the minimum and the overall system could be smarter
dirk: then arrow 6 could be after the NLU
dirk: providers could use their own NLU
jim: what's the difference between the orange (Dialog) NLU and the specific NLU's in the IPA providers
jon: what if IPA provider 1 is American, then it could have full complement of components and functionality.
… we now have text coming from the Dialog box
dirk: could come up with different walkthroughs describing different alternatives
jim: is one of the goals to enable IPA providers to support any input
… are the inputs in step 13 in a standardized format
dirk: 1 -- raw audio, raw ASR or NLU
jim: suggest restricting to one alternative
debbie: I agree
debbie: maybe rule out NLU as a format
dirk: raw audio would be the most common
… or text from a chatbot or ASR
jim: future enhancement would be to standardize NLU, but not covered yet
jim: what is the purpose of the NLU in the Dialog box?
dirk: that's the local NLU that processes some of the queries
Action: dirk to change arrows to reflect the primary use case
output side of walkthrough
dirk: arrow 1 is the various NLU results
dirk: asks dialog registry for the best dialog
dirk: may need to query for missing slots
… gets next dialog move
… NLG comes up with text, doesn't repeat information that's already known to the user
jim: we have a distributed semantic processing, some done in orange and some in blue
dirk: yes
jim: this might be too confusing
dirk: there are some implementations but not standarized
debbie: how do you find the Dialog that goes with the NLU
… if the intent is "book flight" then there could be lots of dialog that could handle that intent
… how do you associate that intent with the appropriate dialog
… so actually intents and entities come back
… and that's mapped to dialogs
dirk: that's correct
debbie: when do you actually get your information from the database
dirk: we haven't gotten to that yet, I could add that
dirk: things need to be clearer
Action: debbie to pull together list of outstanding comments
dirk: also collected some suggestions from a presentation
… what should we do to prove that we did what you describe
jim: eventually there will be conformance tests