in a few me all
there's or a nineteen was for channel i don't necessity but i am i paper
is "'cause" if a chance for adaptive language understanding
this is also my whole
firstly i will give some
with its knowledge about school language understanding
this is the time of spoken dialogue system slu module serves as the interface between
asr
and then
and it reminded management motive
the input of all slu it's word sequence and also is all what is meant
for example the user's is only flies from both the new york
and all for all slu can be
in these to find a flies and the city also
that's it you all you partial is both a and basically also destination is we
off and let the m can make some of these issues about how to give
a good we apply for you
do you recall rate slu can be viewed as a sequence labeling problem that is
included can be what a sequence and the output is a slot sequence
is a example
yes a example the i own representation is used the force higher than that no
it means no slots for a carnival it and eating and i
is a on use the two tasks and you was for long as well
and the and finally we can get us some slot value yes
if we have sufficient in-domain they how with
a human okay she we use easy to us test actually slu a system with
a deep learning models now
you know left part of his
yes em or do for all
one time and the right about everything and then for learning curve all yes em
all at stance that suppose that the performance of the ask him a low heavily
relies on how much data we used for training
us all other all-pole of all probability is that we have no sufficient intimately always
is visual when we need a new domain so a data collection and annotation is
is very it can also
very expensive and time-consuming
so we have to space
then you might result
a small and all the news articles that may or even a totally new dialogue
i will show some examples about the new ideas and use lost
in eliciting to some hosting to train changes it
that's let's say francisco is the city name of stroll okay she
well as the
well with a disgusted to name of a tool okay she and can't afford infinite
maiden name also location in a few times as the data name
location
so i think first and those of the test set all still should
is a relatively new while you to solve a slot or location or the policy
to you "'cause" a low is seen in the training change in their bodies
is not a common to all not at the by from location compensated is doesn't
and viewed as a difference well so expensive and you want but at the seven
isn't as we can find that
i and number
absolute new binding useful
for
probably to some people to layout and mse training data
next we can also classify and a new slots into two
i into a retinue well and absolutely wall
here is for example it does not stop or location or probably can be a
can
can the outcome competition was also you can you can see is the sloth
well applies at least one is that so value one so
so in you all paper we want to tackle the for a relatively new values
and relatively new slots in a conversation with
here we propose one possible way to us also propose all
a relatively new one minus lost
is at one because that we manually speech every slot into a small hands
for at home because that
each other at some concepts
right exactly list in a unified the only one
and a lot of these slots distributed
then as a whole of at all because that
it is for many work for example let's talk about the city of impartial
may have different ways
like
firstly the city name of are located on the city name false at the of
the partial
the speeding that's lost entomology actions we get only one
replacing
one trouble for okay sentence
so the of the partial
last
and then
procedure of using a protocols that is all depends mainly work
do you see is that spatial way case
into now is just one can be represented as a couple of atomic on that
they are some examples of the stuff i don't know about some slots and representations
based on the concept
let's see the user interface also i from because at first but also or a
lot older colours that have a however you are was a relative shaded area
this is an example we have used in the previous slice
if we want to a predicate predict the label also
sometimes
but also is that everything value for that slot for location policy
and the second the contest going to leave is also on the full
but if we model the slots this all at all because that's
we can find that of course there is actually see for
sitting and in the context i'm going to leave
is the or from a location in the chain so let's by the
at one constantly help you
and an overwhelming
we haven't a pretty some new slot at a time
compensation
g is an example if we have only a list two slots a location don't
city name and it worked in boston
maybe also found
the new slot from location tones the name and location on state
g it lists like this about how to morally so that it is not a
because that's
in the traditional model with a rifle because that we have only one class classifier
prediction that's lost
but if we represent
just wanna buy at all
yes or no because that
but strong can be i'll we present
so that i
as a
it impossible for example here of cells at it is defined as a couple of
state and for location
so we propose to simple yes no too much time based on
at home concept
the first method are just there is simply a considers the different part of f
and tahoe as
independent of classification task
g r and of from
yell of data name and from okay she is predicted independent and the by the
way in the i was he might also predicted by and are
another classifier
in a similar mass of the weights useful work considers and a different part of
the at how as
a parallel task
is it you can sample anyway
a lot of trouble pretty if you also location depends on all to
output of fifteen
elliptical least stage
no
the prediction can be can't you declare their collect all atoms in the top or
back here the predicted
and this is a
now what's wrong is represented by a couple of i still
yugoslav maybe produce which is a levels on choice
but we all okay but we didn't
we just should or shouldn't nice
goal slot s and one prediction without any position
a formal nasa a so that only concept
it has been realized khomeini walk
a nice
in this in the nist lots of human knowledge and it may is not so
is there any easy way we want to ask
we define a light
no policy was also or name can be on a sure way for speed and
a slot into single path
right well known and obtain a sequence of surrounding it is very easy and whatnot
well i know what a simple this is not real structure as in the top
of atomic also so
so we propose you will see i think of the model to encode a slot
name into a back to wait for it what no slot surrounding in it is
that can also distribute distributed representation for a small
and first of all we need to make any assumption that is just a name
is that meaning for natural language description
so we will for instance
and last fall subordinating both e
a in this work i didn't ask personally way to have a
a slot encoder is also a yes tomorrow which k
whose input is sorely that we i
no to find a final to the wireless of both i wanna fast and the
forward pass
a concatenating can see that would be a strong looking at
no for distorted we have i story many we drive to our work was actually
at each subordinating and
and there's utilize at a car in the time step
no we can get a scroll it with the same size as data
a smaller number all
and also we k and a softmax normalization is us go back there
no let's go to la experiments we evaluate our method on two task
by the set of mismatch and domain adaptation
the first task but a set of mismatch this all ages
which is widely used it as a benchmark e slu community
it has about five so the centres for change and
my hundred sentences for task
and it is lost this including and every slot is represented by a hubble happened
with that is to the first time energy contours
a first time is introduced for forty five at from the concept and the set
of that is inconsistent atomic
to bidirectional a generalisation july difficult to do for relative regimen use lost
we you and you
task that h is x test which is a mismatch with the changes that you
are mostly want to use some cases about relative a new value
channels
for example
the city name is called a by follow cg and acoustic of the difference in
training in training data
the city name only covered by from c t is relatively new to the slot
to see so we
we just randomly in replacing the while you all to say the in the
it is passed that
without relatively new well known we can data it is x test sample
no
as a challenging
is a experimental results of all the mass at all
it just ages and ages extract
first we can see last let h is x test is really you can challenging
for the traditional so on
so i'll
time model
the performance drops from a ninety five for someone not
and as a as a single best we also add a recognition already feature at
that
additional input for that yes
and it improves
improves
a slightly about one
persons on
or at x k
and bimodal by morally atomic let slot by
at home because that we can find that
let independent model
a unified implement bounded dependent model yet i can actually a you can increase of
the eighties and
and also
it again also case that significant improvement
or a standard at saps
over the original
yes model
as we said that at all because that uses a lot of human operation and
it may be designed so to overcome this weakness and the slot invading us things
be performance
very pointless
us we can find that is a little bit in
you domain and independent
and if we use
you e
you where use i preach everybody may need for initialization we can find that if
they implement it is much problem into a dependable
so fun we'll we also want to a have a look at a low published
result homepages x
in the centre eighties
we have a lightning in if where using a single so that i mean past
our already got with its the past promenade a sickening
we also performed for is german only
but maybe
adaptation to better but it but it in the mocap data for the
you
rate from a new slot
no yes it will use of multiple supply which causes about two thousand dialogues and
yes
is used
i think the target domain but we have you know analysis for
only adaptation
duration was way have several unused possibility is to switch
well the result shows that the training data also must also meant yes two k
helpful an slu model in the target domain
because
nell some slots you can taste initials less all and talking and comments
and it you will model the slot value of s o from concept
the improvement okay a we can get a for it
finally i will give some
two examples
two cases to for discussion that in the left
in left side this about
the
so what confirmed or has infinite learned the svm guitar tone prediction because the requirement
that have something is only for slot
come from
or has tv in the see that it
but the
all more okay
and the actual concept here we use these global or current sample has internet
and read a similar case will rise it is up on a new slot component
or to be a long before never if it is in the selected
after that at times of come from and it should be allowed it is so
all master s or for me
a total conversation okay
can find and yourself
if not i can give some conclusions
and first well defined at all because that's
can have activated at data sparsity problem of slu and method of selecting many which
can be extracted and then
automatic or least in there is very promising it is usual what we also want
to explore popular at some because m with yes
is the whole encourage
maybe we can use some a cross language
what do i
i did try
it was
this example shows two
to slot from cd electricity labels have a part of the city
but in the city name combine these two slots
after different
the and not on the say
so you know training there are low
city names we covered by from c d is housing to reconsider
so if we attack
so the task and if way
you some relatively new
but it was for to set in the test set
if you can be a challenging
for that
to proceed
this model
i see it is very easy because
this design can what my figure
in ages because this we designed a buyer that data from provided
but only for some for something to say that we use we use although for
this paper we use some try
chinese data set we must to
do you do that at one because they have to follow based
for a screen
for each plane labels