0:00:15 | re |
---|
0:00:17 | okay like to welcome everybody see this |
---|
0:00:20 | a special session a natural language generation for dialogue systems |
---|
0:00:24 | i'm just gonna give a five minute overview of the session then we'll have a |
---|
0:00:30 | couple of long talks accomplish very short talks and then we'll have a panel |
---|
0:00:34 | and |
---|
0:00:35 | about a lot of this kind of came together to organize this so the other |
---|
0:00:39 | organisers of their temper like a kind reader david how corrupt |
---|
0:00:43 | showing or read and very in a research |
---|
0:00:47 | so the |
---|
0:00:48 | what the organizers task is to do is to say like why we would have |
---|
0:00:52 | a special |
---|
0:00:54 | session on a nlg for dialogue systems |
---|
0:00:58 | and you know because you might think well this |
---|
0:01:01 | areas been around for a long time indeed some of the earliest work |
---|
0:01:04 | on computational models for natural language generation was work done in the context of a |
---|
0:01:09 | dialogue system or question answering system like kathy make humans |
---|
0:01:13 | early work for phil collins early work |
---|
0:01:17 | there's also a lot of earlier much earlier work you note going back almost twenty |
---|
0:01:22 | years |
---|
0:01:24 | people doing starting to do statistical |
---|
0:01:27 | work on statistical natural language |
---|
0:01:30 | generation |
---|
0:01:31 | starting with the kind of seminal work of like alien like who showed that you |
---|
0:01:35 | could kind of have a very loosely configured hybrid |
---|
0:01:39 | linguistic statistical representation where you could overgenerate and then you could learn |
---|
0:01:44 | rightly rules and filter out |
---|
0:01:47 | this filter out the data's you could produce and so that works as old as |
---|
0:01:51 | nineteen ninety eight so you might |
---|
0:01:52 | still say like why would you have the special session now and also in the |
---|
0:01:57 | context of darpa communicator conversational dialogue systems |
---|
0:02:01 | i edited a special issue with computers |
---|
0:02:04 | speech and language are not sure language generation for spoken dialogue |
---|
0:02:08 | and the number of people that you know had papers in the special issue including |
---|
0:02:12 | am and i lasso and ridge alex rudnicky mari ostendorf stephanie seneff so a lot |
---|
0:02:18 | of long time people been working on conversational dialogue systems |
---|
0:02:22 | for years |
---|
0:02:24 | but the reason that we wanted to have this special session despite the fact that |
---|
0:02:27 | these a lot of that |
---|
0:02:28 | its of generation for dialogue have been around a long time is that there's been |
---|
0:02:33 | a recent kind of resurgence of interest in a natural language generation because of kind |
---|
0:02:40 | of the |
---|
0:02:41 | ai renaissance i guess we should say all the interest in chat bots and all |
---|
0:02:46 | the consumer products that are out there now like collects and google system |
---|
0:02:50 | the other thing is that the availability of large online corpora like open subtitles or |
---|
0:02:57 | twitter or |
---|
0:02:59 | i you know corpora like that have led have may people wonder whether they could |
---|
0:03:03 | actually use a purely |
---|
0:03:04 | statistical kind of and machine translation approach to produce dialogue turns in an open-domain way |
---|
0:03:11 | so there's been a lot of activity in that area for the last five years |
---|
0:03:16 | and |
---|
0:03:18 | so |
---|
0:03:19 | so there's a lot you know seems to be a lot of different stuff going |
---|
0:03:21 | on of this field and one of the reasons why i wanted to organize the |
---|
0:03:25 | special session was the kind of |
---|
0:03:27 | be able to look at what we can do now with different generation techniques into |
---|
0:03:31 | bring you know |
---|
0:03:33 | it especially in the panel to bring in a panel of experts people who worked |
---|
0:03:37 | on one language generation for dialogue systems and try to get some different perspectives |
---|
0:03:42 | on from |
---|
0:03:43 | from their from their points of view what kinds of techniques which ones don't work |
---|
0:03:47 | which things the ready for primetime it could go into consumer products in which things |
---|
0:03:51 | are still just |
---|
0:03:52 | kind of basic |
---|
0:03:55 | research ideas |
---|
0:03:57 | and when am i am particular interest and i think that of many of the |
---|
0:04:00 | other people to organize the panel who have put out these other challenges but e |
---|
0:04:05 | two e challenge is also the web nlg challenge now is in interest and stylistic |
---|
0:04:11 | variation in some of the classic things the natural language generation sabine able to do |
---|
0:04:16 | in their sentence planner and like and a kind of interest in whether |
---|
0:04:21 | and to and framework is actually without a lot of extra architectural details that kind |
---|
0:04:27 | of model the traditional natural language |
---|
0:04:30 | architecture whether they're actually gonna be able to produce different kinds of |
---|
0:04:34 | stylistic variation like the |
---|
0:04:37 | previous generation of statistical language generations could do |
---|
0:04:41 | so that's kind of why were here |
---|
0:04:44 | and |
---|
0:04:46 | we have |
---|
0:04:47 | too long papers |
---|
0:04:49 | redundancy localisation for the conversation of unstructured responses and that a neural language generation in |
---|
0:04:55 | dialogue paper |
---|
0:04:56 | we have to short papers that will be presented in five minutes the for the |
---|
0:05:01 | panel |
---|
0:05:02 | and they'll be in the poster session later i wanted to have them in |
---|
0:05:06 | in the discussion in our in our minds before we start doing the panel |
---|
0:05:11 | because i think |
---|
0:05:13 | there's a really interesting thing here of a new generation challenge that aims to be |
---|
0:05:17 | a little bit more complicated than what people then you seen in the neural generation |
---|
0:05:21 | framework |
---|
0:05:21 | and then somebody was really |
---|
0:05:24 | hot of the starting box |
---|
0:05:27 | and that |
---|
0:05:28 | the corpus was released in two weeks later they had a they this |
---|
0:05:33 | character the character out of the box model that uses the corpus of this is |
---|
0:05:37 | all kind of very much |
---|
0:05:39 | breaking news |
---|
0:05:40 | and |
---|
0:05:42 | so we can go ahead and |
---|
0:05:44 | get started for the for the main papers and then we'll have the panel at |
---|
0:05:48 | the |
---|
0:05:49 | at the end |
---|
0:05:50 | okay so sebastian |
---|