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