and then it means
but not accommodating rumbled university japan
i think i try to talk about alex context kind of ipd strong impression upon
is shown over multiple thereof
and this is done during the lack of all university and the quantity fact leads
to the japan
okay so
she using a dialogue systems and should allow a dyad
and must scotty and that's system should apply a
under a
accumulated knowledge during their ideas design example
i used to a safe you know they are all that and that i-vector line
don't take shown you to pay once for try to write
and another a new that it i know that dialogue i like to fly to
right after the tightest that the other station
a three d the user utterances meetings that systems should apply and this kind of
information about the upright airline pilots
nice
and this can be used for the future recommendation
not correctly this kind of tum-initial is time we prepared by fifteen db don't a
but do you think it should to be applied here and you adding the dialogue
and the we also i'm really a closed domain that talked about
and it's you think know ledgebased if necessary records
assuming the dialogue corpus including all lexical item is unrealistic
so i'll talk about i we want to and make that's talked about
i two i buy a new concept dialogue and that will reduce cost to manually
and knowledge base and now we are building a chat about in the food and
the restaurant domain
a laconic target is to apply two and a
could than a hundred and they worry
and the subset tractable us should be able to continue type dialog even for unknown
talent
the idea example source that you that's it i try to cook now supporting today
and those is that it is not supporting it and on down for that system
and simplest cradle simplest case john is
so what even actually boring but is incredibly meetings such simple hundred maybe abrupt question
if the variable that's
so on a local it is in the lexicon are shown to implicit confirmation and
here that system tried to acquire that on court one protocol cut their worries over
unknown time to form a major
i don't see is also an example of a single example i will try to
class you're going to the and that's a great unknown time interval you
o the system predicted
okay we use one protocol category although this i don't amnesty one
this is done by a british previous method
our previous method
i thought to use these
channel that in the grammar and the four types of japanese a character types
and a
it can really
and this may be a intimacy
note that that's is then generate implicit confirmation request we lacked predictive category c o
for example a good initial restaurant how or not on the geography
and if the user on like i think so
they probably user response
and that extend it only if predicted how they what we see if correct why
not for all a user is this use a response
that system a quick story did i mean that
these categories seem to be to show are quite then that system and a quite
that's not supporting the only two in canadian category
and this and it makes k y ix speech components are all mutually it's not
a good for this task
so
this is mister example of explicit confirmation but
there we need was incorrect a format on be the only one could argue that
system if the model really an italian
so we think that this kind of a comparison we case degraded that you that
extra experience
it's like to hand if also express
confirmation and it is correct
but very well yes so i sometimes have a much more mushroom are not purely
and that system asks if the mushroom are rumoured audio the italian
relative to yes and i think the sets explicitly expression degrade the user experience so
we are now a lot wrote is using a implicit confirmation
okay but so and to determine with that they were either correct or not if
this card
i because you use that the response
you use that is boundaries b is expression
and that includes not only simple affine what people and negative responses
okay these lists either you very edus samples so you that it is picked up
and or yesterday and the that's if their masks i want to one if you
to japanese food
they that users did not exactly well what are you talking about
the by using these as a responsive that system can easily recognise that of the
predicted after what do you want incorrect
on the other hand in it you example or you that it is based upon
goal yesterday under that extensive i wanted to stop and you the food
and they do that
i likely to
so this is a difficult to determine
and italian so
this is not while it does not want a cue to did i mean that
but it is to predict it cut they were collected or not
so a lot of a obvious out a our problem without is to take various
features into consideration and before and after that increase the complementarity with
and i another i
about the other programming to be sort of it in this one that are you
that do not always respond to correctly
so that i don't that means that there are sometimes inconsistent
so this is also incorrect
confirmation via estimate of the onion started around and the japanese with the
and this user response i've just
so
if a guy's if there are you gotta that this
is this indicates that
activity correct
they are incorrect nor its will be added into the system not it's
so i left second problem is old is to exploit responses over multiple dialogues
okay so let me someone it our proposed method is a first one is to
design a deficit
a whole machine learning based classification
and that consider expression i dunno simple of comedy or negative responses
and that miss out that exploits user utterances around i don't know that the mean
before and after
implicit confirmation request
and all i think on the proposed method is to exploit the determination without although
multiple dialogue
and this become possible if other system is deployed on sr for this is a
conventional one-to-one dialogue
but now building but at about and it is the ensemble
so that system channel interact use a multiple you that's
if the same content
so we integrate that is out and
user that without
for determining
without that
quite a pretty story
predicted cutter will easily collect one
okay so this is overview of our method o cost
i've that i think i explained
are you that some unknown town
i do that system generate a implicit confirmation with a
predicted category c and the now use a i thought this so
utterance
and it's and that system calculate the probability p w three from a single user
response at this point i
and the next that system
and according to their responses problem and use that's
that is or like this
then after that we calculate a major role these a probability i'm so by integrating
believe probabilities are be forwarding to find out of confidence major to detach with it
to collect one
okay so this evaluation so i explain to the background on the proposed method and
the problem now i am you explain got a log files may result in more
detail
and the data for extra and experiment
and the next i explained that our
second propose a result
and without on the computer we computed my talk
for five middle part of our proposed with
well
so you calculate the probability that the response is that it is i believe and
without category c for unknown time w if the covariance relative to collect a note
but a
we i introduce our notation for you wanted to
so you don't is therefore user utterance
you are containing the unknown time w
and it's
increase the controversial
group-based
including the predicted included category and you do is that this response to
and the here we use
using logistic regression for
pretty for determining made it is predicted category the correct or not
and we incorporated in table p g s
so for the loop is expressed you do so
not only affirmative or negative expression but also some of the expression and we also
see this you expression and its relationship with what do you wanna under u two
and finally we also incorporated a relationship between you want you to
and to decide are listed and the teachers
so this part of the six
are constant six speech as
under these that
expression in u two
for the for the two is a complex wanted to the baseline and we also
you can incorporate it either voltages
and the second group represent a that express shown you to an adaptive a user
utterance before all have actually correct
and that
the last along its relationship between you and we used that means
are you a way that you want you to contain
the same one and whatnot
and also featured by
what is a before the result hundred
but data collection
so we collected a user utterance it's before and after implicit confirmation request
a fast by of clauses fourteen
the first we ask a walk while "'cause" to encode a think this is really
about a specified by i
for example i eight by how to fold up for that
and so then that system responded
i generate input is to call have initialized an implicit confirmation request
that it is that correct or incorrect
so
these are requests correspond to a this specify the time so we pretty be able
to increase to confirmation requests a for each specified that are
for example italian particle for data twenty well i mean the dishes
and then we ask the user we ask
the walker to respond to do this
a confirmation request
so we pretty the other a twenty channels under their corresponding correct and incorrect if
which to cover image only based
and the we asked
although one hundred workouts
and the quality a lot of two thousand and of their own
and we after that we excluded embodied utterances
what is so this is the result of user logistic regression only ten fold cross
validation
and the we gotta that can cut their policy is correct if the probability was
like larger and larger than zero point five
and of
this low so the baseline this really the proportion result
and this
table shows out a confusion matrix
and we can see that a classification accuracy improved
and
especially no precision of the detection of the product cut a woody
improved
and this the most significant feature was if able that the you might include the
cut they were eating use it is one
so that means and that same topic if the shared what the shared
in
the u one and this one
and that it is you insignificant a feature but not that it's a user included
start of it
then it in this result shows that proposed the p to improve the detection of
incorrect categories
what is needed to move along the next
that second problem in front
for this is a position she are so we take great it's the
probabilities and the
i integrate
that
probabilities
so this continuous major is to determine collect cut they what is wrong in the
user responses
so easy a also used a logistic regression
so we actually we test it as a regression function such as a random forest
additional buttons we showed that it up it out of the logistic regression
i don't we use this by the feature list at each year
and this undercutting what we see a very valid have correct one time with w
a range that computers the major xt does three shows so we change it is
a shorter the value and the if corpus it
exceeds a threshold
the system channel at the same
to that's if they would knowledge base
what is so this is a conditional so we use of that same data
as i explained before
and the we divided them into to rate the training and test with it
to make that experience perfectly open use a block on the request
and the we selected in this policy is happening with that probability a longevity from
a forty nine or forty eight one of the discourse in that we have all
a lot of data fit
and the daily all that in the feature value problem that in response it according
to the computer vision
and this
i really show the result
and this me six we present a fast and that the recognition performance improvement by
multiple user responses the second one is how many sports event need to acquire these
are correctly
they were
the third one is how to fit furniture for constant
well this is that it out for the last question
so we introduced
e but so that the meantime break even point until indicating
it indicates that parting mean precision rate is equal to the recall rate
so i received operational and recall car and the we can see that p b
e p value all and do not have to while larger than
the top and recall while so
that point and the diva any the larger than two
so this means that added to the user response is i had able to improve
the logistic regression functional
performance
and the two on the to determine if the predicted categories collect on
under the second question is how many user responses are needed
so i wrote it
the increase in to be p-value i do that
a horizontal axis is done about in
so this pdf on this graph shows that increases in viterbi peabody while that's right
in the in the one small
so this indicates that it is worthwhile to ask them why users
for
that's a that's what you that's
implicit confirmation request
and the we all we can also see that the deep they diminished in
become between class
so this means that the d needed to be done problem asking what use
the final question is how can we fit to the threshold
so we think that hype original data are required because
that systems should avoid applying incorrect information
so we set high spatial so that the pressure data becomes also almost a one
and the we predict recall rate in this day
so we can see that the recall rate for indy five
well as a zero point two one table one seven five
so it is another all but we think but the old because
we want to avoid a writing imported incorrect information
and we also see that article recall rate but it only increase if we've
so this means that substandard high threshold you know that system to a quite more
categories along with high pressure right
okay so let me somewhat i to this talk
so a lot of timit goal is enabling with a realise that system that allowed
to do that is dialogue
and the tackled in this paper is to determine you stuck at a forty if
correct or not a sort of that you wish to complement your process
and the we propose to the middle part by dividing a feature set and that's
the kind of it integrates the probability o into one complex vector
and the result so that performance was improved
so our future work is o
two for the party line you are used to compare the implicit and explicit confirmation
when you quit
so we assume that an implicit confirmation it's a bit
in that in the viewpoint of the user experience but we need to verify
under the second one is to incorporate the proposed effort into it prototype
the taurus it for your attention
okay so we have about that so many possible questions
and it's
okay
in future work
i
g u i
i
e
i
i think
right
you
three
so you
once
yes it's incentive for your comment and the i think we you
and you want to use it to undo a we need to
carefully designed that
com experiment and
i only that we needed to compare the just techniques speech type and but you
proceed
i and also
explicit and
not a kind of intuitive and based centre for document
a question
so one to the to do so it is to have the system so
is you
rubber goal is to tell you not just cool
so that gives the user chance to say no stop
or just
not do anything so it is not as intrusive so you don't do not problem
but sometimes very clear that this make this assumption
so
one point is don't being with this method in that sense about
so intent to talk about that
one cory to
enjoying the conversation
and you said that is we think that this kind of are repeated based on
you very annoying
four we want to
and on the cheap conversational a continue and it
as a to do that we are introducing a implicit confirmation
you just a i mean i do so it is not a question is just
the state so the user doesn't a response you don't to the dropped it is
you that's
most of your research on
there are lots of the target restaurant
so
that combining the implicit and explicit intuitively if you have a high confidence
understanding you could use the implicit and it very well conference
my is the explicit
you know that seems a little more natural maybe getting over some of the deception
issues
the first question
consider you could also they can
these things got nothing to annoy keeping a threshold how many questions are allowed to
ask but it those kind of techniques
so i already that under a at various you on the data that dialogue strategy
so we just think that repeating with this kind of expressions to computationally very annoying
so but we need to
but i thought you could speechto coverage on implicit confirmation
so i don't
for convenience if one to control that confirmation process
okay so it is just about times so that each sensor speaker