0:00:15so you know everyone my name is german i am up used to in a
0:00:21nice as a fine so today i would like to present my i will we
0:00:27tried to new ways neck to neck which and ways in dialogue using an encoder
0:00:32decoder with a semantic relation
0:00:36so is my present a little bit on the technical know report still based on
0:00:42a don't know four point to sleep okay that's a
0:00:48so i my representation i personally i and is to use some of brief introduction
0:00:55about the not fast so followed by a no general a model for or for
0:01:02the re recording you are then we see that then write the and
0:01:07with we of these and my a mean look my main contribution of in which
0:01:13we is to use a new architecture we called it an ankle do calculate the
0:01:18decoder
0:01:20from the use of forms you agree that so we you're going to prison they
0:01:24have not to kind of a new class and the first part is lively and
0:01:30the second pass and we find a and a lastly we
0:01:34only
0:01:36give some experimental setup and results can lose it
0:01:42so let's start with the introduction of nlg task
0:01:45so is the n is that many
0:01:48i just e
0:01:50okay i don't convert the a meaning representation to an actual
0:01:56and english on a sentence
0:01:57for example we a given but i don't which is a combination of bow
0:02:02the tax and i've we have since i've here for example the inform and the
0:02:07list of slot value pairs for example we have to do a to a to
0:02:13slot value pairs here for someone to new with the value of the hybrid the
0:02:18second one is a poor with c at a value my script
0:02:22so well the generators should and generates the not a sentence to test for example
0:02:30we have a previous is basque restaurant all the second one we had not then
0:02:35with the pirates of last for so what
0:02:39that is to a brief introduction of and as the last
0:02:43so one
0:02:45i believe but you one
0:02:46the
0:02:48new approach based on the
0:02:50on and on a
0:02:52neural architecture so
0:02:55follow from the button to the top given the dialogue act as a pair of
0:03:00dialogue act and system than a to be learned so what the course of a
0:03:06natural language generator is at the lexical i and sentences with a list of example
0:03:13we have three slot names serves as a lot for
0:03:15so for the known choice of open of the acoustic so we can based on
0:03:23the man and
0:03:25i am model lstms you bought some kind of encoder-decoder model so that powerful of
0:03:31c is then we don't use "'em"
0:03:33on also a sentence and after that we have a descendant can be lexical like
0:03:39to perform to form the required sentence
0:03:48so
0:03:49here i and you give museum this general on a general model for the on
0:03:55the record in a new language generator so
0:04:00is an overview of the i and bodies newly a new maybe that generate the
0:04:05which can be divided into two pass the pos on this and go to encode
0:04:11all select and realise how much information it was information
0:04:14the i z the core to the upper side is you could do it uses
0:04:19use the we use the as you can and base model language model
0:04:24so well here is our how many than on a minute contribution in this quilt
0:04:31in ways we a propose a new model colson it and echoed actually very to
0:04:37decode the
0:04:39so
0:04:41yep
0:04:43here's a whole model
0:04:45the
0:04:46and go to know with a new architecture can be divided into a three pass
0:04:52three components so first only and corner to end goal are all you compress the
0:04:56target mini dropped and that a representation
0:04:59this second one is the and then you a new proposed on a component we
0:05:04call to a greater to a lie control the semantic and
0:05:11to refine easier the input sequence
0:05:13and
0:05:15and the c d is you have said one is the decoder used and i
0:05:18and you could or would you see a reply sentences
0:05:22so let's move to a further up to the but so in the decoder side
0:05:27we use the be directly know the are you with and course the separated a
0:05:33parameterization of slots and values
0:05:36and
0:05:38where e the reader consists of two last
0:05:43and a lilac to but and the what was it i don't run light on
0:05:47a representation and the re final to cancun calculated see a new input token able
0:05:55to do the decoder
0:05:57all user is you
0:06:00so let's move to further into the at a later in you know how a
0:06:05model so the a photo one z the lighter calculate c
0:06:10dynamite through a representation with cca a concatenation of slot of accent i
0:06:17and z
0:06:18and as easy the pants and you can isn't based on the absolute value representations
0:06:26and here the refinancing second pass and we finally
0:06:31we'll
0:06:33calculators z new we would x the and put it then i into the
0:06:39and the are using our
0:06:41for the language of then there is an of the sentences
0:06:44and
0:06:45in how what we
0:06:48we further apply the not i don't have a representation to the that you put
0:06:54into the other u c l so firstly the it is you we set and
0:06:59a big it can be a normally five to one
0:07:02to use a on the on
0:07:04as the not i don't know a representation
0:07:07and
0:07:08a sickly z can you did a activation is also modified to depose the influenced
0:07:15by the as you like about a representation
0:07:19so let's move to the we find a so in a in how well a
0:07:24refined as well and examine the choices for the refined of for example we can
0:07:30now we can use the cans and based on see a getting algorithms to apply
0:07:36to the refined a so
0:07:38actually the refine it is highly refined have work week or refinement from seven
0:07:44of cr
0:07:46i don't a web sense it's and dg and the origin to a token that
0:07:51but
0:07:53so for this we use the tense and look as a here is a higher
0:07:59attendant algorithm and is the second one is a getting because the is that the
0:08:06dense and get is them with less and has advanced and apply for the refined
0:08:10of
0:08:10of course the a lighter used so it how what we used a
0:08:15the first attention because it
0:08:17and for the kids and look as and we just apply there's assume a simple
0:08:21one m is wise at least inanimate a multiplication for getting the reply to
0:08:29so that more to further into know how can a
0:08:32how can contrast the not depends in we get them
0:08:36so firstly we just use a simple back to you attend to wait see a
0:08:41see that i
0:08:42and
0:08:44we further we can we'll but with both files that you may be lies in
0:08:49metrics can be back to do
0:08:52another two where the
0:08:54to get information and
0:08:57lastly to not in order to be
0:09:00of course the no
0:09:02put a further than in the context information we propose here to what we can
0:09:09see a pretty as a here
0:09:11a recent i
0:09:13active have a previous know if you story of the to the two of the
0:09:18not dance and
0:09:21so here you we use them getting the guys in which is use the two
0:09:27guy a simple and simple way to study the multiplication and addition
0:09:35so left to the experimental setup
0:09:40is a well
0:09:42we know only
0:09:44we can that the use we only on the
0:09:47under the dataset for model
0:09:50which as the rest of one hotel a laptop and t v
0:09:54we implemented by not using the tensor for all
0:09:58and all of the generators what chain a we see back propagation through time
0:10:03or
0:10:04a stochastic gradient ascent with early stopping we had a l two regularization up to
0:10:11four forty five ginny a examples and the hidden sty dataset is that the c
0:10:18and
0:10:18we said to keep drawing dropped lower rate seventy no for the initialize what is
0:10:25right and position we use a group of a glue
0:10:28and
0:10:30for the evaluation of we use a blues and
0:10:35slot error rate no discourse
0:10:37do you evaluate
0:10:39i will work
0:10:40so
0:10:42here use our result we compare our work with it was intended to form the
0:10:51no
0:10:53which we represent what here and we politically get a got and own something a
0:11:02nice result here we go back to the out the whole models outperform the previous
0:11:07one
0:11:08and
0:11:12use we propose a and because maybe a how work cannot and can varies based
0:11:20on them together and can be varies by on the etsi so what
0:11:24we have is we take i comments and that takes place of five are randomly
0:11:30to supply and that's well
0:11:32so here is the result so use a few go three yes
0:11:39we can see that actually a whole just performs a real someone so in is
0:11:45in the is a peak a three which is a on that so it's beyond
0:11:50by
0:11:53increase mantle step on the other the
0:11:57the a put percent this of training data from what can withstand the training data
0:12:01to the one hundred percent
0:12:05so no the not enough figure four we just a and conduct the a general
0:12:11models in which we pools of we most only the owns it for five i'm
0:12:19not do means and the arch in which a gender issue proposed model
0:12:24now
0:12:24yes on individual domain in here we that's only if it or restore and hotel
0:12:29laptop t v
0:12:32so here is very a little bit nice no is that so is the dense
0:12:38and on the behavior of three models from this we can now sees at how
0:12:46the pole model with the context with second day
0:12:51can
0:12:52can
0:12:54can not as intense in can
0:12:56as a model can at hand and in the am going consecutive of acoustic the
0:13:02tokens
0:13:03so what symbol we can
0:13:05a list of spoken here
0:13:07the phrase okay
0:13:09so here is no with that of the top generated a post from
0:13:14form
0:13:15and no on a
0:13:16compare the from our model with the previous the one syllable
0:13:24so that company would we just and presented no our new model coder and go
0:13:31to a greater decoder in which we z is the with the and can see
0:13:37it up to new but the first one politely used and attention over the input
0:13:41meaning representation
0:13:43the second part is the refined a with the danced and all getting a mechanism
0:13:48to revise the input tokens and on which a model and we can generate or
0:13:55do you see up also model and then we use the evaluation metric a bus
0:14:01and
0:14:02thus
0:14:03score and a slot array as to what you various how well and take us
0:14:07to send
0:14:15thank you we have again
0:14:18six minutes questions
0:14:32i think you're much of the joke
0:14:36can you please maybe i just didn't really
0:14:41see something about the size of your training data and the number of difference
0:14:48to predict looks like compare or liam
0:14:52how much as
0:14:54in addition
0:14:55how many
0:14:59you see a
0:15:00and you replaced all so question
0:15:05it is on a system
0:15:08about
0:15:10initial dataset size
0:15:12how much
0:15:17the image slice
0:15:29which instance
0:15:33again
0:15:39actually i'm sorry the us government and
0:15:44about now five thousand a sentence is the hotel almost seven thousand a lot of
0:15:50and tb is must big with all who in a
0:15:56a route
0:15:57a thirteen a thirteen thousand synthesis
0:16:01that's this aside dataset
0:16:04okay i think the size
0:16:08number of the predicates and model
0:16:12things like compare
0:16:15conditional move the
0:16:17well slowed the legislate
0:16:20one before lost
0:16:23without
0:16:27last slide
0:16:29example low
0:16:31okay sorry
0:16:33so you see the predicted compare
0:16:37and all of the first right
0:16:40compare we aim
0:16:42yes
0:16:44how many things like this predicate are in the mobile
0:16:51alright sites
0:16:55i'm sorry for a not for this for the laptop sale mean
0:16:58well on
0:17:02and
0:17:03and i don't a here is a appear only one time in the dataset
0:17:09for example we can have a man to compare name
0:17:12with the and but we have the same i think i but the is a
0:17:18list of slot value pair is different for example we had in this we have
0:17:22a companion screen size residues and in as a one week have
0:17:28maybe we have compared name is i knew squeezed i also
0:17:33so
0:17:36and not that i don't i hear
0:17:37appear in this latest and laptop and t v just one time so is that
0:17:43stuff up for is such a nice the a model but i have to learn
0:17:50the new one
0:17:51the new the new and
0:17:53i have to learn how to how to
0:17:56applied to the new lose sequence of
0:18:00a slot value pair
0:18:02okay we assume
0:18:13if the input dialogue act with just cry
0:18:17those to a entities what would be the difference in the output
0:18:27or actually in our work here we also a follows that no
0:18:34in the
0:18:37we also follows that uphold okay and up would use the is then routed out
0:18:43as a sentence we the
0:18:46firstly i complete which we can generate owns the of these slot requires a lot
0:18:53for you pass and a six second one also as the is the list of
0:18:58syllable one so well in the country all correct although for example you know
0:19:03in is this one
0:19:09by sampling is one where we have like in input dialogue here
0:19:15so up the incorrect output can be this just output can upload use all the
0:19:26information from the slot value pairs but
0:19:30we can see that the l a own night and the l seventies
0:19:36into in-correct so
0:19:39but not
0:19:39as it problem but you know in our model we can now and generates the
0:19:44do you at the correct all of the week lies a reply
0:19:50plentiful of the that say for example in that example you just had
0:19:54what if the screen sizes range with the same value for both entities
0:19:59so they both have a large would you then leave large out because it's not
0:20:05it's not different so there's no comparison
0:20:08actually in how well the g different value of screen
0:20:15is not is not problem it but
0:20:19also we came back here
0:20:23the same the same value is not know
0:20:25about
0:20:26the value of slot followed by us a lot and of a span of
0:20:32so the value of slot is not know
0:20:35a
0:20:36is not important in it is cool because
0:20:40in here example
0:20:42we have at least of celebrity by so we
0:20:46not really skyline the and as the
0:20:50i understand that you delexicalise but a human we do something different if the slot
0:20:54values of the same versus a slot values are different
0:20:57so it doesn't make sense if you comparing two things and they both have the
0:21:01same slot value
0:21:03that is it makes sense to say
0:21:05just say this one's large and batman's large
0:21:10instead of say for example they both are large or not mention that slot at
0:21:15all because that's the same value for both one
0:21:19how well actually is exactly z is the a post processing when
0:21:23which is
0:21:24and
0:21:25system is an for the first he
0:21:30just a the post processing after we have a and i was no candidates the
0:21:38that is collect a sentence and we lexical i descendant to inform the oars in
0:21:45the one
0:21:46so that's why insist that
0:21:57so thank you very
0:22:00i presenter again