0:00:14thank you
0:00:16my name is madonna kernel or on the proteins to depend
0:00:19i'm going to
0:00:21put in this talk with this title
0:00:24well
0:00:26i'm sorry this try to us it's almost everything but it me have been out
0:00:32into me
0:00:33for detecting in
0:00:35vol the goal overall result is to build in the real data systems that use
0:00:41that are willing to use
0:00:43why we focus on interview data fifteen i because they can be used for collecting
0:00:48information from humans and
0:00:51they can organise that you permission
0:00:54and users are expected to the scroll zero a parser information to want to welcome
0:01:00eighty seconds on the human system
0:01:03and
0:01:04although below
0:01:07interviewed items a commercial potential
0:01:11well quite we focus on systems that use of the willing to use a because
0:01:17some applications need to be used repeatedly all of
0:01:22for example systems will die recording and the decoding is able to have been a
0:01:28i mean
0:01:29need to be used d v d three
0:01:34their couple previous one
0:01:36though
0:01:38you have you need not a popular applications all the time system
0:01:44a database source but there are couple a fist and i mean that have been
0:01:50very useful for a defensible would have the same
0:01:54or rating current scroll the as this temple government pencils are assigned surveys and a
0:02:02simple five at all people's use about the future role
0:02:06so it's us this then focus
0:02:10manual obtaining likely than much information from users and
0:02:15i'm you're not sure people are willing to use these distinctive utt
0:02:21well our approach e
0:02:23the twenty minute interview dialogue system
0:02:26and
0:02:27codebook
0:02:30because that's what to carol users to enjoy a conversation
0:02:34and also had there a couple a meter well all other studies
0:02:40that's shows more talking of use increases the user look up on pins on an
0:02:45engagement
0:02:46and
0:02:48some studies
0:02:50so that i want to increase the price
0:02:56or
0:02:58there are two possible approaches to integrate well i need to deal with that of
0:03:03the primary strategy and the sometimes in books
0:03:06interview and the second is the to deal with interview that the primary sort of
0:03:12the end sometime thing using
0:03:14in that smalltalk the proposed approach needs closer to human conversations but it might can
0:03:21into in many an utterance is because that the current technology all utterances can we
0:03:26do not good that i
0:03:28you mode
0:03:30and the second approach is not are about right you have the advantage
0:03:34that
0:03:36it can go back to the interview you meet of multiple
0:03:40wrong
0:03:41so we think that second approach
0:03:46we'll only implement each other
0:03:49face and based on our approach it is a japanese text based interview data system
0:03:55for direct recording
0:03:57it asks the user all other heroes the day before
0:04:03on the like the comedies and like this
0:04:06all other systems that what did you have the what we proceed the data and
0:04:10i and i have here
0:04:13and that smalltalk
0:04:15starts directly
0:04:19well the objective of this system is to hold in rocky information on all of
0:04:24what the user up on each key
0:04:27you know they can use the of the user directly having not he did and
0:04:32that it at all
0:04:34the computation time t or i dunno
0:04:38time
0:04:41the simple
0:04:43knowing that you by three
0:04:48and this is architectural
0:04:50and then explain each more do
0:04:54the first analysis and all that if you got a mute point was on all
0:04:58japanese well known that known
0:05:02if you can mute one was sensible japanese and not
0:05:06nothing the well
0:05:09multiple other company
0:05:12well
0:05:14countries okay for a system also note that three hundred and the ball and things
0:05:20and
0:05:21their approach finding a fruit groups a corresponding to meet accomplishments chi psi the maintenance
0:05:29one this new on form
0:05:33and the language understanding problem
0:05:38i press creation and semantic content
0:05:41extraction that address
0:05:43contracts creation
0:05:45classify the user utterance and the three types screen in and negative out and then
0:05:51the only thing about that
0:05:52and the number all utterance type is more because the interview data you a bit
0:05:58of thing
0:06:00simple
0:06:01and
0:06:02we try to a system based on need classified of a comedy about
0:06:08and we use logistic regression trees probable words pete rose to all qualities classification
0:06:15and the semantic content extraction on a extract five kinds of information namely food and
0:06:21drink in reading loop amount would and i'm having good
0:06:26and we use
0:06:27they are very high will be talking missile then the dictionary lookup
0:06:32and the training data can to crawl up by
0:06:36a fifty six hundred out of it
0:06:41and
0:06:45and i mean want to and dialogue management role in my view
0:06:50all
0:06:52e p the frame based dialogue management
0:06:54and
0:06:57the prince we like i
0:07:00well let us assume that there is sixteen like p and the user utterance e
0:07:07but
0:07:08understood and the type you can i the system
0:07:13phone lines the type is the from team on the content you like this
0:07:18and
0:07:21a knowledge of ac to happen that and each user found
0:07:25to be one this mean and that's anti use put here
0:07:30and based on this claim a the next
0:07:35system based on
0:07:36like to have anything at all rounds used in it
0:07:43in anaemic screen
0:07:45could group operation
0:07:47four point and extracted you mean if not in an utterance
0:07:52is that system needs to know each could group because and the it needs to
0:07:59know
0:08:01you
0:08:03peering the frame that the with the name should be
0:08:07and
0:08:10well that system utterance i while you're right it's could groups will narrow using would
0:08:16go a middle
0:08:18that's one
0:08:19like the
0:08:22well
0:08:23now in its roles it the system estimates the to the group using e
0:08:29the name and thus could the names sorted on features using the logistic regression and
0:08:34generate
0:08:35in articulation
0:08:38like this so this is a binary system
0:08:43the
0:08:44in need determined based on a video probably you but i don't mean that a
0:08:49detailed explanation for that
0:08:54in from all gotten in a joint ugly
0:09:00like the
0:09:02well that's
0:09:03candidates for the system smalltalk utterances are selected from a predefined a four hundred what
0:09:09into account
0:09:11based on the type and the content all preceding you that
0:09:15opening
0:09:17for example when the user utterance is a problem of p and negative a more
0:09:21utterance there already
0:09:24and
0:09:25the useful forty two whatever
0:09:29were created using the based on your are on
0:09:35we have something
0:09:37but you also model got currently like
0:09:41these it is my favourite fruit you know example you scroll and great need showing
0:09:48input the and do you write sampling you know example asking a creation
0:09:57finally item explain a about direction at that stage some order to is one utterance
0:10:03from i liked and it's from
0:10:05or which sentence mortal apparently
0:10:09and
0:10:11me how a very i we do very simple strategy
0:10:16and the number all the important thing a mortal
0:10:21after each user only right system based on needs fixed to n
0:10:27so
0:10:29in times of extremes in an exchange it's of course after a
0:10:35and coral to the information
0:10:38i and each star a small talk after it's randomly
0:10:43children from county
0:10:49we conducted a user study to
0:10:54investigate the effectiveness of the a small talk about it
0:10:59well we compare the three constant the first one is no used to you condition
0:11:07that mean unique or their no other words the number of this cannot be in
0:11:12each mortal you that's on all available
0:11:16at this is the baseline
0:11:19we also compare one is to use on the john and three is the condition
0:11:26okay that's we use the you condition
0:11:30mean the number of this came out that the in each one two three
0:11:39we recorded it one hundred participants by a problem solving
0:11:45and we didn't collect there are also provide function in the on each i mean
0:11:51for the and they don't have to they've but a you know you
0:11:57and
0:12:00that the participants are is that i don't talk about to engage in a be
0:12:04the fist enemy the three conditions then the overall content or not
0:12:10after it's better they were asked to evaluate that of it in table writing on
0:12:15a five a point you
0:12:19the much analysis didn't answer
0:12:21a limited to seventy three to avoid too long a conversation
0:12:28i
0:12:30well we what it what have a tuple or hundred but this one
0:12:36but we found that a partition of the dialog albeit party on
0:12:44programs that's that the
0:12:48and it's not a matter liking that in writing
0:12:51or else is a know how the program
0:12:56well we use the
0:12:57on the data or one in nine into participant
0:13:03a basis in and like this
0:13:07of course of noise you can be shown on the these normal in-car
0:13:12the language understanding of home and
0:13:17like is utterance type classification accuracy nine the one point important and semantic wanting extracts
0:13:25accuracy is the whole point
0:13:29okay well then you don't know
0:13:31bad
0:13:32and also the anybody could group estimation accuracy
0:13:37but you for when the robot in that
0:13:42this is not this is also you know
0:13:47okay
0:13:47these right examples all correctly dialogues
0:13:52or noise you only john and one is you on john and
0:13:56three st you condition
0:14:00one if you only on dial
0:14:05i don't have more or
0:14:08shown in a rate one and o
0:14:12also in three is the you only on dial
0:14:16longer or
0:14:17that's model we use forty
0:14:26and
0:14:27is
0:14:28a sort and showed in user input is shown
0:14:35well okay a related problem
0:14:39it was
0:14:41sort the scroll saw noise do you ones on and a blue well it's all
0:14:45the
0:14:46scores for one is to you only some and three
0:14:51agreement balls so that
0:14:55scroll three is the you condition
0:14:58in
0:15:02it
0:15:03e
0:15:04so of course last simplicity
0:15:08a for simplicity and noise the u is the based because we there's no one
0:15:14score
0:15:17and we found that no one is you brought down
0:15:23noise you cornerstone your
0:15:27there are
0:15:31i zero it aims at like one and what do you want to talk in
0:15:37and library
0:15:41and we also found that a three d you
0:15:45is a good
0:15:49well is to you
0:15:51all zero i in like naturalness
0:15:55want to talking and i've renice
0:15:59although
0:16:00and the
0:16:01there are no
0:16:03statistical significance
0:16:05well work want to talk again and library in it but
0:16:10the average
0:16:12all so us to you is worse than one is t
0:16:19well
0:16:22then we discuss this one
0:16:24impatient roles
0:16:25three is the you are not a good at one is to you reading this
0:16:30is probably big wheel including the number of local content
0:16:35ladies the possibly yield in it and that are not than this
0:16:40because of the probably you know
0:16:44generating
0:16:50appropriate model gotta
0:16:54me a problem by
0:16:58in the upper an appropriate initial model that the
0:17:05well we don't only
0:17:08so
0:17:10alogue all three of this to you condition
0:17:13and but you in buying deletion
0:17:17and w found that i
0:17:20at to pretend all the process have a small talk about it so
0:17:26appropriate but
0:17:28only twenty eight percent also explored a small talk about the
0:17:35i appropriate
0:17:37to e
0:17:40and that's why
0:17:42it's est you
0:17:45no not given a good patient
0:17:50in
0:17:50and maybe conclude
0:17:53this goal
0:17:57at home if h is a like you proposed to denny
0:18:02modal got are used to improve user input is shown all we have used an
0:18:06existing
0:18:07and
0:18:09the recorded over user study using a japanese text based interview dialogues example
0:18:15that the recording shortstop smalltalk utterances eva blowing pressure on to the user
0:18:23it is also so this did start anything too many small talk utterances make
0:18:29makes the user's impression words people are they greedy you want to be and it
0:18:39it increases the possibility of anything learned to a better
0:18:45well any future problem to a on the another is a buddy green bay you
0:18:52how can anything small talk about that a fixed the gpu you
0:18:56use of the system
0:18:58or maybe
0:18:59and will eat
0:19:01they were all systems that you problem waiting to use repeatedly
0:19:06but
0:19:08the
0:19:10user study reported in
0:19:12these paul well you mean because the use that system twenty one
0:19:18so
0:19:20we need to investigate the issue
0:19:26in another study
0:19:28another is applied
0:19:29and
0:19:31and on a peaceful future work is to the robot missile role is a broader
0:19:38direction on this phone or what
0:19:40okay and the number all smalltalk on the fixed
0:19:45but i think that you
0:19:49is important to me than our number or also mortal thoughts it's great several and
0:19:54depending on the appropriate can is so the generated
0:20:00i mean a small talk about that and you're are currently working on
0:20:07thank you very much
0:21:04you understand a quick the based on he that i
0:21:10well the
0:21:11that to see smalltalk utterances
0:21:14from the predefined with it
0:21:21and three d are using a very
0:21:24i mean
0:21:25a simple
0:21:30immediately it is simple
0:21:36risk like this self training that wine the is that policies upfront even always the
0:21:42negative odyssey
0:21:43but
0:21:44mm
0:21:46all of course you know that at least a simple and
0:21:51you need to do well on a corpus based missile two
0:21:55ginit each
0:21:58o appropriate that's multiple utterances and three me are trying to use various the using
0:22:07various features like about a not only about the words but also
0:22:13i mean
0:22:17type of utterances and that of history and e p
0:22:23a about enough amount of data
0:22:27maybe we you
0:22:30us to use
0:22:32i mean deep running our here is the in based is able to
0:22:38to the one multi
0:22:42to a more most appropriate utterances of based on a dialogue context
0:23:38so you have the statistics showing how the frequency of acceptable smalltalk remarks decreased as
0:23:47you had second and third remarks and that seemed like a
0:23:52possible explanation for why people prefer the one with one verses three utterances but i
0:24:00am wondering if you
0:24:02have the possibility to look at just the subset of cases that had more than
0:24:08one acceptable remark and looking to see whether that had a had a different behavior
0:24:13from the overall set of
0:24:16three smalltalk utterances
0:24:23you mean
0:24:26if
0:24:27what happens if
0:24:29all three
0:24:31about that is actually a right well we haven't sixty that
0:24:38o
0:24:44probably an excuse to look at and
0:24:50so to divide the i mean
0:24:56okay
0:24:57but
0:24:58sorry
0:25:00but
0:25:02dialogue
0:25:03all the
0:25:06also each time with very long and that there are many a small talk about
0:25:10things and rory all
0:25:16all its mortal
0:25:17in one utterance in
0:25:20all objects or there are
0:25:23the sound quality works we are and some well doesn't or where
0:25:28but this might be a good possibility for a following experiment specifically looking at good
0:25:33versus not so good multi
0:25:35i think it
0:25:38it's good to know the user feel about for each column huh
0:25:44by asking the another approach found to rate