Speech Transcript - Concept Transfer Learning for Adaptive Language Understanding

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

Concept Transfer Learning for Adaptive Language Understanding

Oral Session 5: State Tracking

Su Zhu and Kai Yu