0:00:15 | okay so |
---|
0:00:16 | hi everyone |
---|
0:00:18 | and one's a and i would like to |
---|
0:00:19 | talk about the new data sets that you can't in a |
---|
0:00:25 | that i nine we have |
---|
0:00:27 | created at three at what university and it's a dataset design for |
---|
0:00:34 | and two and the natural language generation |
---|
0:00:36 | with that we mean that a generating a fully from data and from unaligned |
---|
0:00:45 | data pairs so that means a pair of the meaning representation and the corresponding textual |
---|
0:00:49 | reference |
---|
0:00:51 | with no water additional annotation |
---|
0:00:54 | this has already been down but so far all the approach is where limited to |
---|
0:01:00 | relatively small datasets and all of them use of delexicalization |
---|
0:01:06 | induce are the datasets you can see on the slide |
---|
0:01:09 | and our goal here is to go a bit for the with the data driven |
---|
0:01:14 | approach and to replicate the |
---|
0:01:17 | rich dialogue can discourse phenomena |
---|
0:01:20 | that had been targeted but you know the year and non end-to-end the rule based |
---|
0:01:24 | or also statistical approaches |
---|
0:01:28 | and what |
---|
0:01:28 | we have down is |
---|
0:01:31 | we have collected a new training dataset that should be challenging enough to |
---|
0:01:37 | show |
---|
0:01:38 | some |
---|
0:01:40 | more interesting outputs more interesting sentences |
---|
0:01:43 | and |
---|
0:01:44 | it is also much bigger than all the previous datasets we have over fifty thousand |
---|
0:01:49 | pairs of meaning representations and textual references |
---|
0:01:55 | the textual references a longer so we usually |
---|
0:01:58 | have more sentences that's |
---|
0:02:00 | describe |
---|
0:02:01 | one meaning representation and the sentences themselves are all also longer than in previous datasets |
---|
0:02:07 | we |
---|
0:02:08 | have also made the effort to collect the data set in as divers way as |
---|
0:02:13 | possible |
---|
0:02:14 | and that's why we used editorial |
---|
0:02:18 | instructions to crowd workers on a |
---|
0:02:21 | crowdsourcing website |
---|
0:02:23 | and |
---|
0:02:24 | we have found out that this leads to more divers descriptions so |
---|
0:02:29 | if you if you look at these two examples |
---|
0:02:33 | you we have a low cost |
---|
0:02:35 | japanese-style cuisine and |
---|
0:02:36 | you we have cheap japanese food so the |
---|
0:02:39 | descriptions are very diapers and |
---|
0:02:42 | also there's more of them on average than in most previous nlg datasets we have |
---|
0:02:48 | more than eight |
---|
0:02:50 | our preference texts better meaning representation |
---|
0:02:55 | we have evaluated the dataset in various ways and compared it with the previous datasets |
---|
0:03:02 | in the same domain |
---|
0:03:03 | and we have found out that |
---|
0:03:05 | we have |
---|
0:03:08 | higher lexical richness which means |
---|
0:03:11 | more |
---|
0:03:12 | divers text and terms of words used and a higher proportion of rare words in |
---|
0:03:19 | the data |
---|
0:03:20 | the sentences are also |
---|
0:03:23 | on average more syntactically complex so we have |
---|
0:03:29 | longer and more complex sentences |
---|
0:03:32 | and we have also up |
---|
0:03:34 | us |
---|
0:03:35 | kind of a semantic challenge because we asked the crowd workers only to verbalise information |
---|
0:03:40 | that seems relevant given the instructional picture so actually this requires content selection also for |
---|
0:03:49 | natural language generation which hasn't notes it's not present in the previous |
---|
0:03:55 | state of sets of the same type |
---|
0:03:58 | and we are organising a shell challenge with this dataset so |
---|
0:04:03 | you can |
---|
0:04:05 | all register for the challenge we would like to encourage you to do so |
---|
0:04:09 | and try to train your own nlg system and |
---|
0:04:15 | sub made your results |
---|
0:04:16 | by the end of local october |
---|
0:04:18 | we provide the data and also a baseline system along with the baseline system outputs |
---|
0:04:23 | and metrics creates |
---|
0:04:25 | is that |
---|
0:04:26 | will be used for the challenge along with us some human evolution |
---|
0:04:32 | so |
---|
0:04:33 | is it and i woods |
---|
0:04:35 | like to invite you to comment c or a poster later on and we can |
---|
0:04:40 | talk about this some more |
---|
0:04:42 | and definitely |
---|
0:04:44 | and downloads the data and take part in your challenge |
---|
0:04:48 | thank you |
---|