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