0:00:15 | i have become a nested into |
---|
0:00:18 | today i will see a presentation on this topic and be compatible dependent you can |
---|
0:00:25 | buy jones for the difficulty |
---|
0:00:28 | i'm cutting question and a domain but that it out in director who have role |
---|
0:00:32 | and gt |
---|
0:00:36 | at fast and |
---|
0:00:38 | i'm trying to show up background to about this study |
---|
0:00:44 | they the gusto disability or |
---|
0:00:47 | mann and thompson or information in it may be because |
---|
0:00:52 | you could be built into the discourse structure as a question is that a tree |
---|
0:00:56 | structure |
---|
0:00:58 | and generate a rst discourse telling about data you below by a strong how to |
---|
0:01:04 | be used in many in can be applications |
---|
0:01:07 | because three thirty two for my and computationally intractable |
---|
0:01:13 | for example i was but have be used for automatic discourse parsing and they some |
---|
0:01:21 | additional autocorrelation |
---|
0:01:24 | and so on |
---|
0:01:27 | on the other hand and we simply some of these chats |
---|
0:01:31 | have proposed another strategy that the discourse given basically |
---|
0:01:38 | they planned event initiate more about the question she |
---|
0:01:42 | in some cases you |
---|
0:01:45 | it's pdas shows |
---|
0:01:47 | that is that it's all text summarization for something they are well-known methods |
---|
0:01:54 | including markers |
---|
0:01:56 | text summarization method |
---|
0:01:59 | a u g |
---|
0:02:01 | at the at data |
---|
0:02:05 | also propose a pity summarization mister based on typically shady coding and |
---|
0:02:11 | he got |
---|
0:02:12 | hi point |
---|
0:02:14 | i don't |
---|
0:02:15 | marcu and other |
---|
0:02:18 | the that's but that |
---|
0:02:22 | dependency based discourse syntax have used a potential to improve the modeling or |
---|
0:02:29 | a wide range it'll take the based applications |
---|
0:02:34 | how it but unfortunately the |
---|
0:02:38 | it is no large scale dtd whole but in english |
---|
0:02:43 | little research it's hard try to create e d compact |
---|
0:02:49 | and i think i you know and the at and propose to on but |
---|
0:02:55 | r is the gtd a two d dct that |
---|
0:03:00 | in this lack a we compare that to combine them |
---|
0:03:05 | i love a seasonal comparing the combined you are mr a as follows |
---|
0:03:11 | after that we make the two missiles algorithm |
---|
0:03:15 | if i to create |
---|
0:03:17 | it can |
---|
0:03:19 | a which format this complex |
---|
0:03:22 | room that depend basically asked would be point of view |
---|
0:03:25 | that which format that and i more accurately by automatic have that |
---|
0:03:33 | finally which comments almost equal to critics summarization |
---|
0:03:41 | i an example |
---|
0:03:43 | rst discuss the |
---|
0:03:45 | you know |
---|
0:03:47 | e one e to it and corresponds to elementary use cost you need in that |
---|
0:03:54 | at |
---|
0:03:58 | any valuing excluded by curly brackets in the same sentence |
---|
0:04:05 | i'm just saying |
---|
0:04:07 | a domain but are going to buy a rhetorical relation and of course |
---|
0:04:14 | identity of discourse unit |
---|
0:04:17 | each discourse unit can see so i don't cool s |
---|
0:04:23 | and i mean |
---|
0:04:23 | and you go yes they reach |
---|
0:04:26 | expresses more important information down |
---|
0:04:31 | on the other hand is |
---|
0:04:33 | means that started i state each added within |
---|
0:04:37 | some fourteen information |
---|
0:04:42 | according variations are also separate it into two types |
---|
0:04:47 | one any more nuclear and you have a multinuclear |
---|
0:04:53 | mallets union has only a multinuclear relation e |
---|
0:04:58 | one and the one s is close to any while a multinuclear relation pinks is |
---|
0:05:06 | that there are no discourse you |
---|
0:05:10 | the recovery shown that can be also glue to use a detailed class q that |
---|
0:05:17 | share it we gotta meaning such as and i've already shown that say nearly thin |
---|
0:05:23 | part of a conditional and so on |
---|
0:05:27 | actually i will not try to home but d r s t e d structure |
---|
0:05:33 | into discourse dependency tree structure |
---|
0:05:37 | also next i really to go into the t k to all their combined |
---|
0:05:45 | well for one you will hear you conditions |
---|
0:05:49 | is |
---|
0:05:49 | that's you know entity |
---|
0:05:52 | methods have at the same combine |
---|
0:05:56 | i just example quality so already more nuclei detection |
---|
0:06:05 | you manipulation each you can see so and i don't and |
---|
0:06:12 | and a because any more important element the s |
---|
0:06:17 | so |
---|
0:06:18 | you know anybody could yes you wanted to |
---|
0:06:23 | in this really show a |
---|
0:06:27 | it belongs to be equal |
---|
0:06:30 | at the right |
---|
0:06:34 | all other education models you don't two d because the guy |
---|
0:06:42 | the next the lead to a she you don't to e |
---|
0:06:48 | and to the next higher data |
---|
0:06:52 | you don't do |
---|
0:06:56 | and finally a |
---|
0:06:59 | according if you don't have to be and to be a bad the |
---|
0:07:06 | to the aperture |
---|
0:07:09 | if |
---|
0:07:12 | and iris we discuss three |
---|
0:07:14 | consider so only multinuclear relations |
---|
0:07:18 | takes you out and v mister produce the same dependency graph |
---|
0:07:25 | how about four one multinuclear relation |
---|
0:07:28 | you know and d |
---|
0:07:30 | a use different way to combat |
---|
0:07:36 | at least example has brown and multinuclear relation |
---|
0:07:44 | you skiing |
---|
0:07:47 | that if the most and he's a |
---|
0:07:49 | mostly in both that |
---|
0:07:51 | quality the most important |
---|
0:07:53 | and the remaining this three d |
---|
0:07:56 | you got |
---|
0:08:00 | in this case the and e |
---|
0:08:03 | a building a you can see the old |
---|
0:08:07 | to see |
---|
0:08:19 | the next either a |
---|
0:08:21 | a middle to see because if we just sat there i |
---|
0:08:27 | and the to the next higher level or a c belongs to be |
---|
0:08:33 | and finally at the attached to the partition |
---|
0:08:39 | i one the other hand |
---|
0:08:40 | easy house cleaning |
---|
0:08:43 | or end of the three d |
---|
0:08:45 | and you want |
---|
0:08:47 | and four or c t and that's to the an excellent very many |
---|
0:08:55 | the next iteration more new variational |
---|
0:09:02 | the next door neighbour manipulation you have a question the |
---|
0:09:06 | it's almost and |
---|
0:09:08 | the us |
---|
0:09:10 | the most in speech and the main a and it could you know to be |
---|
0:09:15 | for all x p long to see |
---|
0:09:19 | a because she the leftmost nucleus |
---|
0:09:25 | in an image which haven't had variation but at that the |
---|
0:09:31 | is higher than in any and or c |
---|
0:09:34 | the anti |
---|
0:09:36 | you are going to e |
---|
0:09:39 | finally need you know the active shuddered |
---|
0:09:44 | now we can see the different either be you know and be sixteen |
---|
0:09:50 | and you can see the most important difference you that you know in a complete |
---|
0:09:55 | you see many crossing act |
---|
0:10:01 | that's you know methods have another serious program |
---|
0:10:06 | by most sentence si |
---|
0:10:09 | for example be strapped rst discourse the |
---|
0:10:14 | but like you know metal all possible with the simple beauty restructure but then test |
---|
0:10:21 | in this example |
---|
0:10:23 | the sentence |
---|
0:10:25 | consisting or used and eighty eight |
---|
0:10:28 | e nineteen |
---|
0:10:30 | had to |
---|
0:10:32 | collected at that same as you we always the student boutiques about tell the |
---|
0:10:41 | to solve this program used by the proposed a post editing missile |
---|
0:10:47 | pitch attitude the |
---|
0:10:50 | you we extend the lexical model the one |
---|
0:10:54 | that is the most |
---|
0:10:57 | you think example that do you need you nine |
---|
0:11:02 | you can be attached |
---|
0:11:03 | two in the next the most into uni you see it |
---|
0:11:10 | it see that very simple updating |
---|
0:11:13 | but as shown in the next that i used simple was it being you see |
---|
0:11:20 | the complexity or that's reality deep |
---|
0:11:23 | format the rest carry |
---|
0:11:29 | now we can have got you there are two methods |
---|
0:11:34 | you may know that we combat its three hundred to eighty five r is the |
---|
0:11:40 | dg in this is cool dependence you |
---|
0:11:46 | last week and it got still |
---|
0:11:49 | the difference is that you max happening in the number of no leaving the text |
---|
0:11:55 | it is well on the whole |
---|
0:11:59 | max patterning need the maximum number four |
---|
0:12:04 | act on that a scroll on i picture due to some belief no |
---|
0:12:09 | not me that |
---|
0:12:10 | number one or leaving that x |
---|
0:12:13 | x |
---|
0:12:16 | roamed easily that it is could that |
---|
0:12:19 | there you know dtd format e |
---|
0:12:21 | what role and channel |
---|
0:12:24 | a fairly format |
---|
0:12:28 | and that we can see that the you for it that's in but it |
---|
0:12:34 | the large gap between be a |
---|
0:12:38 | here we use more complex change goal |
---|
0:12:41 | costing tax |
---|
0:12:42 | e d u |
---|
0:12:45 | you can see that you or weight project e |
---|
0:12:49 | a you have a wide there is no question e |
---|
0:12:52 | he is added |
---|
0:12:55 | why are you know but for not have many cross syntax |
---|
0:13:00 | okay but i'll a proposal was in that are generally use many costing that scroll |
---|
0:13:06 | on the else |
---|
0:13:07 | but |
---|
0:13:11 | x |
---|
0:13:12 | and we get a good use their own automatic discourse dependency parsing |
---|
0:13:19 | we compared to have a |
---|
0:13:23 | one d in this paper that |
---|
0:13:25 | respectively provide you see |
---|
0:13:28 | and on predicting and distinctive from a user |
---|
0:13:34 | that at a charity are |
---|
0:13:37 | and which produce cheese |
---|
0:13:38 | and r is the discuss t for my but can and then be compactly |
---|
0:13:44 | you do you see that each combined you will |
---|
0:13:55 | problem use table we have assumed that we suppose that |
---|
0:13:59 | so any usually you did it directly or it can but it wrong |
---|
0:14:06 | i do you see |
---|
0:14:07 | because to me |
---|
0:14:11 | i have a |
---|
0:14:14 | you got |
---|
0:14:18 | i post us sixty six point |
---|
0:14:22 | and you know got low a |
---|
0:14:26 | fifty five point |
---|
0:14:30 | well of accurately db them up on break down the tree format |
---|
0:14:35 | i have a period of time |
---|
0:14:38 | fe |
---|
0:14:39 | which you do it that we call this is directed at different |
---|
0:14:45 | next we conduct experiments on text summarization |
---|
0:14:49 | to compare the different dependencies that yet |
---|
0:14:57 | be used to |
---|
0:14:59 | you know text summarization network based on |
---|
0:15:02 | for any all i'll a better discourse the game quickly |
---|
0:15:07 | retrieved while the current faded mister |
---|
0:15:11 | it's text summarization moral probability i do you have some program and |
---|
0:15:17 | if you migration can be stored but in figuring out programming |
---|
0:15:23 | for the details pretty that you fight tutees with your paper |
---|
0:15:30 | but at a whole month or text summarization |
---|
0:15:34 | fatty documents |
---|
0:15:36 | which |
---|
0:15:39 | it man made of course online |
---|
0:15:43 | i that |
---|
0:15:45 | but for do d t |
---|
0:15:48 | you know you did you hallmark got it high boring |
---|
0:15:53 | the other formats |
---|
0:15:55 | but the difference is not the last |
---|
0:16:00 | the hunt and when you want a dtd |
---|
0:16:06 | also go high sampling |
---|
0:16:08 | in the different see that |
---|
0:16:13 | well that it indeed the whole paper is there a g e |
---|
0:16:19 | combating from i d e |
---|
0:16:21 | is it because you know dvds over the in is in this paper |
---|
0:16:27 | what religion channel so important element that you really |
---|
0:16:34 | to active channel two and it can be ut extracting back to be mapped to |
---|
0:16:39 | a program |
---|
0:16:41 | i do you see that the wrong at advantage or you know was d v |
---|
0:16:44 | d format |
---|
0:16:46 | or but we did you format |
---|
0:16:51 | in calculating used presentation we need to do so |
---|
0:16:56 | some of the key point |
---|
0:16:59 | used by the compared to combine from |
---|
0:17:03 | r is the dt in the pdtb |
---|
0:17:07 | well you make that i would be if it is clear |
---|
0:17:12 | and the |
---|
0:17:15 | well experimentally that we show that |
---|
0:17:19 | but for mapping what are correct target |
---|
0:17:23 | i'll about for text summarization you know the method has come at the boundaries |
---|
0:17:29 | although mister |
---|
0:17:33 | if you tell that we need you the two digit format score of the task |
---|
0:17:39 | of white sample and discourse |
---|
0:17:41 | three banks by dividing mister doctor adding |
---|
0:17:45 | and the classifier evaluation |
---|
0:17:48 | in you to laugh |
---|
0:17:49 | we did not use |
---|
0:17:52 | maybe a colleague are not there are so each feature |
---|
0:17:56 | so it would you experiments using a particular |
---|
0:18:00 | we plan to investigate the |
---|
0:18:03 | two d for my if is used it |
---|
0:18:06 | adorable my presentation and give a metaphor |
---|
0:18:40 | a offline and which one and two |
---|
0:18:46 | apply |
---|
0:18:47 | including all of these two but |
---|
0:19:25 | no impact know it |
---|
0:19:27 | really |
---|
0:19:31 | something |
---|
0:19:55 | i g |
---|
0:19:57 | that |
---|
0:20:02 | actually |
---|
0:20:08 | i'm going |
---|
0:20:30 | but in a |
---|
0:20:33 | two |
---|
0:20:34 | vol |
---|
0:20:38 | easy to be a high enough that okay you |
---|