Re: Incremental recognition, Unobtrusive response

David,

   Yes, partial recognition results could probably be merged into a 
single category with external events.  That's not how the V3 draft is 
written (external events really must come from outside the VXML 
interpreter), but we could define a category of asynchronous events that 
could be internally or externally generated.

As for how disruptive these asynchronous events would be, I think that 
would primarily depend on how much we were willing to modify the VXML 
control model (a.k.a. the FIA) to be more flexible.

You are right that a more sophisticated language model could probably 
handle context shifts during recognition.  We should also think of what 
else might change as the result of the unobtrusive feedback.  For 
example, might we want to change the volume of prompts while they are 
playing?  Right now VXML is designed to "package up" a conversational 
turn - prompts and grammar - and to run them as a block.   With 
unobtrusive feedback, as with asynchronous events, the model needs to 
become more flexible.

There are really two separate questions here:  what we think the right 
model is for more responsive user interaction, and how we would have to 
re-write VoiceXML to implement such it.

- Jim


On 11/17/2016 4:39 PM, David Pautler wrote:
>
> Thanks for sharing that info about the proposed external events 
> module, Jim. When the draft mentions how disruptive external events 
> might be to a flow, I would guess most of that complication would 
> arise for catch's outside of a grammar element. Does that sound right?
>
> After reading that draft, it seems to me that Dirk's partial tag might 
> be subsumed by a catch of an 'externalevent' with an inner 'if' to 
> check what the event is, and whether other recent events merit an 
> (unobtrusive) response. And if the response doesn't require jumping 
> out of the active grammar, we could continue getting partial 
> interpretations this way.
>
> Jim, you also rightly point out that unobtrusive responses slightly 
> alter the context for the user and may therefore require a change in 
> grammar. I suspect, though, that language models could be trained with 
> optional changes in context midway through training examples. What do 
> you think?
>
> Thank you, Deborah, for sharing the Semaine project; I hadn't heard of 
> it. I'm planning to do a comparative analysis of the three projects 
> for Dirk's doc.
>
> Cheers,
> David
>

Received on Thursday, 17 November 2016 22:07:53 UTC