0:00:16and then it means
0:00:17but not accommodating rumbled university japan
0:00:20i think i try to talk about alex context kind of ipd strong impression upon
0:00:24is shown over multiple thereof
0:00:27and this is done during the lack of all university and the quantity fact leads
0:00:31to the japan
0:00:34okay so
0:00:35she using a dialogue systems and should allow a dyad
0:00:40and must scotty and that's system should apply a
0:00:44under a
0:00:45accumulated knowledge during their ideas design example
0:00:49i used to a safe you know they are all that and that i-vector line
0:00:53don't take shown you to pay once for try to write
0:00:57and another a new that it i know that dialogue i like to fly to
0:01:02right after the tightest that the other station
0:01:05a three d the user utterances meetings that systems should apply and this kind of
0:01:11information about the upright airline pilots
0:01:15nice
0:01:16and this can be used for the future recommendation
0:01:21not correctly this kind of tum-initial is time we prepared by fifteen db don't a
0:01:27but do you think it should to be applied here and you adding the dialogue
0:01:32and the we also i'm really a closed domain that talked about
0:01:36and it's you think know ledgebased if necessary records
0:01:40assuming the dialogue corpus including all lexical item is unrealistic
0:01:46so i'll talk about i we want to and make that's talked about
0:01:52i two i buy a new concept dialogue and that will reduce cost to manually
0:01:59and knowledge base and now we are building a chat about in the food and
0:02:05the restaurant domain
0:02:06a laconic target is to apply two and a
0:02:10could than a hundred and they worry
0:02:13and the subset tractable us should be able to continue type dialog even for unknown
0:02:18talent
0:02:20the idea example source that you that's it i try to cook now supporting today
0:02:26and those is that it is not supporting it and on down for that system
0:02:31and simplest cradle simplest case john is
0:02:35so what even actually boring but is incredibly meetings such simple hundred maybe abrupt question
0:02:42if the variable that's
0:02:44so on a local it is in the lexicon are shown to implicit confirmation and
0:02:50here that system tried to acquire that on court one protocol cut their worries over
0:02:56unknown time to form a major
0:02:59i don't see is also an example of a single example i will try to
0:03:04class you're going to the and that's a great unknown time interval you
0:03:09o the system predicted
0:03:12okay we use one protocol category although this i don't amnesty one
0:03:18this is done by a british previous method
0:03:23our previous method
0:03:25i thought to use these
0:03:28channel that in the grammar and the four types of japanese a character types
0:03:33and a
0:03:35it can really
0:03:37and this may be a intimacy
0:03:40note that that's is then generate implicit confirmation request we lacked predictive category c o
0:03:48for example a good initial restaurant how or not on the geography
0:03:53and if the user on like i think so
0:03:58they probably user response
0:04:00and that extend it only if predicted how they what we see if correct why
0:04:06not for all a user is this use a response
0:04:11that system a quick story did i mean that
0:04:16these categories seem to be to show are quite then that system and a quite
0:04:21that's not supporting the only two in canadian category
0:04:26and this and it makes k y ix speech components are all mutually it's not
0:04:31a good for this task
0:04:34so
0:04:35this is mister example of explicit confirmation but
0:04:40there we need was incorrect a format on be the only one could argue that
0:04:45system if the model really an italian
0:04:48so we think that this kind of a comparison we case degraded that you that
0:04:53extra experience
0:04:55it's like to hand if also express
0:04:58confirmation and it is correct
0:05:01but very well yes so i sometimes have a much more mushroom are not purely
0:05:07and that system asks if the mushroom are rumoured audio the italian
0:05:11relative to yes and i think the sets explicitly expression degrade the user experience so
0:05:18we are now a lot wrote is using a implicit confirmation
0:05:25okay but so and to determine with that they were either correct or not if
0:05:31this card
0:05:33i because you use that the response
0:05:35you use that is boundaries b is expression
0:05:38and that includes not only simple affine what people and negative responses
0:05:43okay these lists either you very edus samples so you that it is picked up
0:05:48and or yesterday and the that's if their masks i want to one if you
0:05:52to japanese food
0:05:54they that users did not exactly well what are you talking about
0:05:58the by using these as a responsive that system can easily recognise that of the
0:06:03predicted after what do you want incorrect
0:06:06on the other hand in it you example or you that it is based upon
0:06:10goal yesterday under that extensive i wanted to stop and you the food
0:06:14and they do that
0:06:16i likely to
0:06:18so this is a difficult to determine
0:06:21and italian so
0:06:23this is not while it does not want a cue to did i mean that
0:06:28but it is to predict it cut they were collected or not
0:06:31so a lot of a obvious out a our problem without is to take various
0:06:37features into consideration and before and after that increase the complementarity with
0:06:43and i another i
0:06:45about the other programming to be sort of it in this one that are you
0:06:49that do not always respond to correctly
0:06:53so that i don't that means that there are sometimes inconsistent
0:06:58so this is also incorrect
0:07:01confirmation via estimate of the onion started around and the japanese with the
0:07:06and this user response i've just
0:07:10so
0:07:12if a guy's if there are you gotta that this
0:07:15is this indicates that
0:07:20activity correct
0:07:21they are incorrect nor its will be added into the system not it's
0:07:27so i left second problem is old is to exploit responses over multiple dialogues
0:07:35okay so let me someone it our proposed method is a first one is to
0:07:39design a deficit
0:07:41a whole machine learning based classification
0:07:44and that consider expression i dunno simple of comedy or negative responses
0:07:51and that miss out that exploits user utterances around i don't know that the mean
0:07:57before and after
0:07:58implicit confirmation request
0:08:01and all i think on the proposed method is to exploit the determination without although
0:08:06multiple dialogue
0:08:07and this become possible if other system is deployed on sr for this is a
0:08:14conventional one-to-one dialogue
0:08:16but now building but at about and it is the ensemble
0:08:20so that system channel interact use a multiple you that's
0:08:26if the same content
0:08:27so we integrate that is out and
0:08:30user that without
0:08:34for determining
0:08:36without that
0:08:37quite a pretty story
0:08:38predicted cutter will easily collect one
0:08:42okay so this is overview of our method o cost
0:08:46i've that i think i explained
0:08:48are you that some unknown town
0:08:51i do that system generate a implicit confirmation with a
0:08:56predicted category c and the now use a i thought this so
0:09:01utterance
0:09:02and it's and that system calculate the probability p w three from a single user
0:09:07response at this point i
0:09:10and the next that system
0:09:13and according to their responses problem and use that's
0:09:18that is or like this
0:09:20then after that we calculate a major role these a probability i'm so by integrating
0:09:29believe probabilities are be forwarding to find out of confidence major to detach with it
0:09:35to collect one
0:09:38okay so this evaluation so i explain to the background on the proposed method and
0:09:43the problem now i am you explain got a log files may result in more
0:09:48detail
0:09:49and the data for extra and experiment
0:09:51and the next i explained that our
0:09:54second propose a result
0:09:56and without on the computer we computed my talk
0:10:01for five middle part of our proposed with
0:10:04well
0:10:06so you calculate the probability that the response is that it is i believe and
0:10:11without category c for unknown time w if the covariance relative to collect a note
0:10:16but a
0:10:18we i introduce our notation for you wanted to
0:10:24so you don't is therefore user utterance
0:10:28you are containing the unknown time w
0:10:31and it's
0:10:34increase the controversial
0:10:36group-based
0:10:37including the predicted included category and you do is that this response to
0:10:43and the here we use
0:10:45using logistic regression for
0:10:50pretty for determining made it is predicted category the correct or not
0:10:56and we incorporated in table p g s
0:11:00so for the loop is expressed you do so
0:11:06not only affirmative or negative expression but also some of the expression and we also
0:11:13see this you expression and its relationship with what do you wanna under u two
0:11:20and finally we also incorporated a relationship between you want you to
0:11:26and to decide are listed and the teachers
0:11:30so this part of the six
0:11:33are constant six speech as
0:11:35under these that
0:11:38expression in u two
0:11:40for the for the two is a complex wanted to the baseline and we also
0:11:46you can incorporate it either voltages
0:11:50and the second group represent a that express shown you to an adaptive a user
0:11:56utterance before all have actually correct
0:11:58and that
0:12:00the last along its relationship between you and we used that means
0:12:04are you a way that you want you to contain
0:12:08the same one and whatnot
0:12:10and also featured by
0:12:14what is a before the result hundred
0:12:17but data collection
0:12:19so we collected a user utterance it's before and after implicit confirmation request
0:12:25a fast by of clauses fourteen
0:12:27the first we ask a walk while "'cause" to encode a think this is really
0:12:33about a specified by i
0:12:37for example i eight by how to fold up for that
0:12:41and so then that system responded
0:12:45i generate input is to call have initialized an implicit confirmation request
0:12:51that it is that correct or incorrect
0:12:54so
0:12:56these are requests correspond to a this specify the time so we pretty be able
0:13:01to increase to confirmation requests a for each specified that are
0:13:07for example italian particle for data twenty well i mean the dishes
0:13:13and then we ask the user we ask
0:13:17the walker to respond to do this
0:13:19a confirmation request
0:13:22so we pretty the other a twenty channels under their corresponding correct and incorrect if
0:13:28which to cover image only based
0:13:30and the we asked
0:13:32although one hundred workouts
0:13:34and the quality a lot of two thousand and of their own
0:13:38and we after that we excluded embodied utterances
0:13:44what is so this is the result of user logistic regression only ten fold cross
0:13:49validation
0:13:51and the we gotta that can cut their policy is correct if the probability was
0:13:57like larger and larger than zero point five
0:14:01and of
0:14:02this low so the baseline this really the proportion result
0:14:06and this
0:14:07table shows out a confusion matrix
0:14:10and we can see that a classification accuracy improved
0:14:14and
0:14:17especially no precision of the detection of the product cut a woody
0:14:24improved
0:14:25and this the most significant feature was if able that the you might include the
0:14:31cut they were eating use it is one
0:14:33so that means and that same topic if the shared what the shared
0:14:38in
0:14:39the u one and this one
0:14:42and that it is you insignificant a feature but not that it's a user included
0:14:48start of it
0:14:50then it in this result shows that proposed the p to improve the detection of
0:14:54incorrect categories
0:14:57what is needed to move along the next
0:14:59that second problem in front
0:15:04for this is a position she are so we take great it's the
0:15:10probabilities and the
0:15:12i integrate
0:15:13that
0:15:14probabilities
0:15:17so this continuous major is to determine collect cut they what is wrong in the
0:15:21user responses
0:15:23so easy a also used a logistic regression
0:15:27so we actually we test it as a regression function such as a random forest
0:15:32additional buttons we showed that it up it out of the logistic regression
0:15:38i don't we use this by the feature list at each year
0:15:42and this undercutting what we see a very valid have correct one time with w
0:15:47a range that computers the major xt does three shows so we change it is
0:15:51a shorter the value and the if corpus it
0:15:57exceeds a threshold
0:15:59the system channel at the same
0:16:01to that's if they would knowledge base
0:16:05what is so this is a conditional so we use of that same data
0:16:10as i explained before
0:16:12and the we divided them into to rate the training and test with it
0:16:17to make that experience perfectly open use a block on the request
0:16:22and the we selected in this policy is happening with that probability a longevity from
0:16:28a forty nine or forty eight one of the discourse in that we have all
0:16:33a lot of data fit
0:16:34and the daily all that in the feature value problem that in response it according
0:16:39to the computer vision
0:16:42and this
0:16:44i really show the result
0:16:46and this me six we present a fast and that the recognition performance improvement by
0:16:52multiple user responses the second one is how many sports event need to acquire these
0:16:57are correctly
0:16:58they were
0:16:59the third one is how to fit furniture for constant
0:17:04well this is that it out for the last question
0:17:08so we introduced
0:17:12e but so that the meantime break even point until indicating
0:17:16it indicates that parting mean precision rate is equal to the recall rate
0:17:22so i received operational and recall car and the we can see that p b
0:17:28e p value all and do not have to while larger than
0:17:35the top and recall while so
0:17:39that point and the diva any the larger than two
0:17:43so this means that added to the user response is i had able to improve
0:17:48the logistic regression functional
0:17:51performance
0:17:52and the two on the to determine if the predicted categories collect on
0:17:59under the second question is how many user responses are needed
0:18:03so i wrote it
0:18:07the increase in to be p-value i do that
0:18:11a horizontal axis is done about in
0:18:15so this pdf on this graph shows that increases in viterbi peabody while that's right
0:18:22in the in the one small
0:18:24so this indicates that it is worthwhile to ask them why users
0:18:29for
0:18:30that's a that's what you that's
0:18:33implicit confirmation request
0:18:36and the we all we can also see that the deep they diminished in
0:18:41become between class
0:18:43so this means that the d needed to be done problem asking what use
0:18:49the final question is how can we fit to the threshold
0:18:52so we think that hype original data are required because
0:18:57that systems should avoid applying incorrect information
0:19:01so we set high spatial so that the pressure data becomes also almost a one
0:19:06and the we predict recall rate in this day
0:19:11so we can see that the recall rate for indy five
0:19:15well as a zero point two one table one seven five
0:19:20so it is another all but we think but the old because
0:19:25we want to avoid a writing imported incorrect information
0:19:30and we also see that article recall rate but it only increase if we've
0:19:35so this means that substandard high threshold you know that system to a quite more
0:19:40categories along with high pressure right
0:19:44okay so let me somewhat i to this talk
0:19:47so a lot of timit goal is enabling with a realise that system that allowed
0:19:54to do that is dialogue
0:19:56and the tackled in this paper is to determine you stuck at a forty if
0:19:59correct or not a sort of that you wish to complement your process
0:20:04and the we propose to the middle part by dividing a feature set and that's
0:20:09the kind of it integrates the probability o into one complex vector
0:20:14and the result so that performance was improved
0:20:19so our future work is o
0:20:21two for the party line you are used to compare the implicit and explicit confirmation
0:20:27when you quit
0:20:28so we assume that an implicit confirmation it's a bit
0:20:33in that in the viewpoint of the user experience but we need to verify
0:20:38under the second one is to incorporate the proposed effort into it prototype
0:20:43the taurus it for your attention
0:20:51okay so we have about that so many possible questions
0:21:04and it's
0:21:08okay
0:21:10in future work
0:21:12i
0:21:17g u i
0:21:24i
0:21:25e
0:21:33i
0:21:45i think
0:21:50right
0:21:54you
0:22:03three
0:22:06so you
0:22:09once
0:22:18yes it's incentive for your comment and the i think we you
0:22:22and you want to use it to undo a we need to
0:22:28carefully designed that
0:22:30com experiment and
0:22:32i only that we needed to compare the just techniques speech type and but you
0:22:37proceed
0:22:38i and also
0:22:40explicit and
0:22:42not a kind of intuitive and based centre for document
0:22:47a question
0:22:49so one to the to do so it is to have the system so
0:22:55is you
0:22:58rubber goal is to tell you not just cool
0:23:01so that gives the user chance to say no stop
0:23:05or just
0:23:06not do anything so it is not as intrusive so you don't do not problem
0:23:11but sometimes very clear that this make this assumption
0:23:15so
0:23:16one point is don't being with this method in that sense about
0:23:22so intent to talk about that
0:23:26one cory to
0:23:28enjoying the conversation
0:23:30and you said that is we think that this kind of are repeated based on
0:23:35you very annoying
0:23:36four we want to
0:23:39and on the cheap conversational a continue and it
0:23:43as a to do that we are introducing a implicit confirmation
0:23:49you just a i mean i do so it is not a question is just
0:23:52the state so the user doesn't a response you don't to the dropped it is
0:23:58you that's
0:23:59most of your research on
0:24:01there are lots of the target restaurant
0:24:20so
0:24:22that combining the implicit and explicit intuitively if you have a high confidence
0:24:29understanding you could use the implicit and it very well conference
0:24:34my is the explicit
0:24:38you know that seems a little more natural maybe getting over some of the deception
0:24:41issues
0:24:43the first question
0:24:44consider you could also they can
0:24:46these things got nothing to annoy keeping a threshold how many questions are allowed to
0:24:50ask but it those kind of techniques
0:24:55so i already that under a at various you on the data that dialogue strategy
0:25:02so we just think that repeating with this kind of expressions to computationally very annoying
0:25:07so but we need to
0:25:10but i thought you could speechto coverage on implicit confirmation
0:25:15so i don't
0:25:17for convenience if one to control that confirmation process
0:25:27okay so it is just about times so that each sensor speaker