0:00:15in a few me all
0:00:18there's or a nineteen was for channel i don't necessity but i am i paper
0:00:23is "'cause" if a chance for adaptive language understanding
0:00:27this is also my whole
0:00:29firstly i will give some
0:00:31with its knowledge about school language understanding
0:00:35this is the time of spoken dialogue system slu module serves as the interface between
0:00:42and then
0:00:43and it reminded management motive
0:00:46the input of all slu it's word sequence and also is all what is meant
0:00:50for example the user's is only flies from both the new york
0:00:56and all for all slu can be
0:00:58in these to find a flies and the city also
0:01:03that's it you all you partial is both a and basically also destination is we
0:01:08off and let the m can make some of these issues about how to give
0:01:13a good we apply for you
0:01:18do you recall rate slu can be viewed as a sequence labeling problem that is
0:01:23included can be what a sequence and the output is a slot sequence
0:01:29is a example
0:01:35yes a example the i own representation is used the force higher than that no
0:01:42it means no slots for a carnival it and eating and i
0:01:47is a on use the two tasks and you was for long as well
0:01:54and the and finally we can get us some slot value yes
0:02:02if we have sufficient in-domain they how with
0:02:04a human okay she we use easy to us test actually slu a system with
0:02:12a deep learning models now
0:02:16you know left part of his
0:02:19yes em or do for all
0:02:22one time and the right about everything and then for learning curve all yes em
0:02:27all at stance that suppose that the performance of the ask him a low heavily
0:02:34relies on how much data we used for training
0:02:38us all other all-pole of all probability is that we have no sufficient intimately always
0:02:46is visual when we need a new domain so a data collection and annotation is
0:02:53is very it can also
0:02:55very expensive and time-consuming
0:02:58so we have to space
0:03:00then you might result
0:03:01a small and all the news articles that may or even a totally new dialogue
0:03:09i will show some examples about the new ideas and use lost
0:03:15in eliciting to some hosting to train changes it
0:03:19that's let's say francisco is the city name of stroll okay she
0:03:23well as the
0:03:26well with a disgusted to name of a tool okay she and can't afford infinite
0:03:31maiden name also location in a few times as the data name
0:03:37so i think first and those of the test set all still should
0:03:42is a relatively new while you to solve a slot or location or the policy
0:03:47to you "'cause" a low is seen in the training change in their bodies
0:03:53is not a common to all not at the by from location compensated is doesn't
0:03:58and viewed as a difference well so expensive and you want but at the seven
0:04:03isn't as we can find that
0:04:05i and number
0:04:08absolute new binding useful
0:04:14probably to some people to layout and mse training data
0:04:20next we can also classify and a new slots into two
0:04:24i into a retinue well and absolutely wall
0:04:28here is for example it does not stop or location or probably can be a
0:04:34can the outcome competition was also you can you can see is the sloth
0:04:39well applies at least one is that so value one so
0:04:45so in you all paper we want to tackle the for a relatively new values
0:04:52and relatively new slots in a conversation with
0:04:58here we propose one possible way to us also propose all
0:05:02a relatively new one minus lost
0:05:04is at one because that we manually speech every slot into a small hands
0:05:11for at home because that
0:05:13each other at some concepts
0:05:16right exactly list in a unified the only one
0:05:20and a lot of these slots distributed
0:05:23then as a whole of at all because that
0:05:27it is for many work for example let's talk about the city of impartial
0:05:33may have different ways
0:05:36firstly the city name of are located on the city name false at the of
0:05:41the partial
0:05:43the speeding that's lost entomology actions we get only one
0:05:48one trouble for okay sentence
0:05:52so the of the partial
0:05:57and then
0:05:58procedure of using a protocols that is all depends mainly work
0:06:04do you see is that spatial way case
0:06:10into now is just one can be represented as a couple of atomic on that
0:06:14they are some examples of the stuff i don't know about some slots and representations
0:06:21based on the concept
0:06:25let's see the user interface also i from because at first but also or a
0:06:31lot older colours that have a however you are was a relative shaded area
0:06:38this is an example we have used in the previous slice
0:06:42if we want to a predicate predict the label also
0:06:48but also is that everything value for that slot for location policy
0:06:53and the second the contest going to leave is also on the full
0:07:00but if we model the slots this all at all because that's
0:07:05we can find that of course there is actually see for
0:07:08sitting and in the context i'm going to leave
0:07:11is the or from a location in the chain so let's by the
0:07:17at one constantly help you
0:07:20and an overwhelming
0:07:23we haven't a pretty some new slot at a time
0:07:30g is an example if we have only a list two slots a location don't
0:07:35city name and it worked in boston
0:07:37maybe also found
0:07:39the new slot from location tones the name and location on state
0:07:46g it lists like this about how to morally so that it is not a
0:07:52because that's
0:07:54in the traditional model with a rifle because that we have only one class classifier
0:08:01prediction that's lost
0:08:05but if we represent
0:08:08just wanna buy at all
0:08:10yes or no because that
0:08:11but strong can be i'll we present
0:08:14so that i
0:08:15as a
0:08:17it impossible for example here of cells at it is defined as a couple of
0:08:21state and for location
0:08:23so we propose to simple yes no too much time based on
0:08:29at home concept
0:08:31the first method are just there is simply a considers the different part of f
0:08:37and tahoe as
0:08:40independent of classification task
0:08:42g r and of from
0:08:45yell of data name and from okay she is predicted independent and the by the
0:08:51way in the i was he might also predicted by and are
0:08:56another classifier
0:09:00in a similar mass of the weights useful work considers and a different part of
0:09:04the at how as
0:09:07a parallel task
0:09:10is it you can sample anyway
0:09:13a lot of trouble pretty if you also location depends on all to
0:09:18output of fifteen
0:09:22elliptical least stage
0:09:27the prediction can be can't you declare their collect all atoms in the top or
0:09:34back here the predicted
0:09:38and this is a
0:09:39now what's wrong is represented by a couple of i still
0:09:46yugoslav maybe produce which is a levels on choice
0:09:49but we all okay but we didn't
0:09:51we just should or shouldn't nice
0:09:55goal slot s and one prediction without any position
0:10:02a formal nasa a so that only concept
0:10:07it has been realized khomeini walk
0:10:10a nice
0:10:12in this in the nist lots of human knowledge and it may is not so
0:10:17is there any easy way we want to ask
0:10:21we define a light
0:10:22no policy was also or name can be on a sure way for speed and
0:10:26a slot into single path
0:10:29right well known and obtain a sequence of surrounding it is very easy and whatnot
0:10:37well i know what a simple this is not real structure as in the top
0:10:41of atomic also so
0:10:44so we propose you will see i think of the model to encode a slot
0:10:48name into a back to wait for it what no slot surrounding in it is
0:10:54that can also distribute distributed representation for a small
0:11:01and first of all we need to make any assumption that is just a name
0:11:05is that meaning for natural language description
0:11:08so we will for instance
0:11:11and last fall subordinating both e
0:11:16a in this work i didn't ask personally way to have a
0:11:21a slot encoder is also a yes tomorrow which k
0:11:25whose input is sorely that we i
0:11:29no to find a final to the wireless of both i wanna fast and the
0:11:34forward pass
0:11:35a concatenating can see that would be a strong looking at
0:11:41no for distorted we have i story many we drive to our work was actually
0:11:46at each subordinating and
0:11:51and there's utilize at a car in the time step
0:11:54no we can get a scroll it with the same size as data
0:11:58a smaller number all
0:12:01and also we k and a softmax normalization is us go back there
0:12:09no let's go to la experiments we evaluate our method on two task
0:12:15by the set of mismatch and domain adaptation
0:12:19the first task but a set of mismatch this all ages
0:12:23which is widely used it as a benchmark e slu community
0:12:29it has about five so the centres for change and
0:12:34my hundred sentences for task
0:12:37and it is lost this including and every slot is represented by a hubble happened
0:12:43with that is to the first time energy contours
0:12:48a first time is introduced for forty five at from the concept and the set
0:12:52of that is inconsistent atomic
0:12:55to bidirectional a generalisation july difficult to do for relative regimen use lost
0:13:02we you and you
0:13:04task that h is x test which is a mismatch with the changes that you
0:13:09are mostly want to use some cases about relative a new value
0:13:16for example
0:13:19the city name is called a by follow cg and acoustic of the difference in
0:13:25training in training data
0:13:27the city name only covered by from c t is relatively new to the slot
0:13:32to see so we
0:13:34we just randomly in replacing the while you all to say the in the
0:13:40it is passed that
0:13:42without relatively new well known we can data it is x test sample
0:13:48as a challenging
0:13:51is a experimental results of all the mass at all
0:13:56it just ages and ages extract
0:14:00first we can see last let h is x test is really you can challenging
0:14:07for the traditional so on
0:14:08so i'll
0:14:10time model
0:14:11the performance drops from a ninety five for someone not
0:14:16and as a as a single best we also add a recognition already feature at
0:14:23additional input for that yes
0:14:25and it improves
0:14:29a slightly about one
0:14:32persons on
0:14:35or at x k
0:14:37and bimodal by morally atomic let slot by
0:14:43at home because that we can find that
0:14:45let independent model
0:14:47a unified implement bounded dependent model yet i can actually a you can increase of
0:14:54the eighties and
0:14:56and also
0:15:01it again also case that significant improvement
0:15:04or a standard at saps
0:15:06over the original
0:15:08yes model
0:15:12as we said that at all because that uses a lot of human operation and
0:15:16it may be designed so to overcome this weakness and the slot invading us things
0:15:22be performance
0:15:24very pointless
0:15:26us we can find that is a little bit in
0:15:30you domain and independent
0:15:32and if we use
0:15:35you e
0:15:36you where use i preach everybody may need for initialization we can find that if
0:15:41they implement it is much problem into a dependable
0:15:48so fun we'll we also want to a have a look at a low published
0:15:52result homepages x
0:15:54in the centre eighties
0:15:57we have a lightning in if where using a single so that i mean past
0:16:03our already got with its the past promenade a sickening
0:16:09we also performed for is german only
0:16:12but maybe
0:16:13adaptation to better but it but it in the mocap data for the
0:16:20rate from a new slot
0:16:23no yes it will use of multiple supply which causes about two thousand dialogues and
0:16:30is used
0:16:32i think the target domain but we have you know analysis for
0:16:37only adaptation
0:16:40duration was way have several unused possibility is to switch
0:16:48well the result shows that the training data also must also meant yes two k
0:16:53helpful an slu model in the target domain
0:16:59nell some slots you can taste initials less all and talking and comments
0:17:07and it you will model the slot value of s o from concept
0:17:11the improvement okay a we can get a for it
0:17:17finally i will give some
0:17:18two examples
0:17:20two cases to for discussion that in the left
0:17:24in left side this about
0:17:28so what confirmed or has infinite learned the svm guitar tone prediction because the requirement
0:17:33that have something is only for slot
0:17:36come from
0:17:37or has tv in the see that it
0:17:40but the
0:17:42all more okay
0:17:44and the actual concept here we use these global or current sample has internet
0:17:50and read a similar case will rise it is up on a new slot component
0:17:56or to be a long before never if it is in the selected
0:18:01after that at times of come from and it should be allowed it is so
0:18:05all master s or for me
0:18:08a total conversation okay
0:18:10can find and yourself
0:18:13if not i can give some conclusions
0:18:16and first well defined at all because that's
0:18:18can have activated at data sparsity problem of slu and method of selecting many which
0:18:24can be extracted and then
0:18:26automatic or least in there is very promising it is usual what we also want
0:18:31to explore popular at some because m with yes
0:19:01is the whole encourage
0:19:11maybe we can use some a cross language
0:19:15what do i
0:19:19i did try
0:19:52it was
0:19:56this example shows two
0:19:58to slot from cd electricity labels have a part of the city
0:20:02but in the city name combine these two slots
0:20:05after different
0:20:07the and not on the say
0:20:09so you know training there are low
0:20:13city names we covered by from c d is housing to reconsider
0:20:17so if we attack
0:20:20so the task and if way
0:20:23you some relatively new
0:20:25but it was for to set in the test set
0:20:29if you can be a challenging
0:20:30for that
0:20:32to proceed
0:20:33this model
0:21:06i see it is very easy because
0:21:10this design can what my figure
0:21:15in ages because this we designed a buyer that data from provided
0:21:25but only for some for something to say that we use we use although for
0:21:30this paper we use some try
0:21:33chinese data set we must to
0:21:38do you do that at one because they have to follow based
0:21:42for a screen
0:21:44for each plane labels