WEBVTT

00:00:00.800 --> 00:00:03.233
<c.white>Hi, I’m Marc and I’d like to present</c>

00:00:03.233 --> 00:00:07.099
<c.white>how the DGS Corpus implements ethical open data practices</c>

00:00:07.099 --> 00:00:09.533
<c.white>such as the CARE and FAIR principles.</c>

00:00:09.766 --> 00:00:12.666
<c.white>My co-author for this presentation is Thomas Hanke,</c>

00:00:12.666 --> 00:00:17.066
<c.white>but the work I describe has been contributed to by many people over the years,</c>

00:00:17.066 --> 00:00:21.300
<c.white>so credit should go to the entire DGS-Korpus project team.</c>

00:00:26.600 --> 00:00:30.233
<c.white>A quick note for those who are not familiar with signed languages,</c>

00:00:30.233 --> 00:00:33.766
<c.white>DGS stands for “Deutsche Gebärdensprache”, </c>

00:00:33.766 --> 00:00:35.633
<c.white>meaning German Sign Language,</c>

00:00:35.633 --> 00:00:39.600
<c.white>which is a signed language that is primarily used in Germany and Luxembourg.</c>

00:00:45.833 --> 00:00:48.600
<c.white>As is the case for basically all signed languages,</c>

00:00:48.600 --> 00:00:51.066
<c.white>DGS is severely under-resourced</c>

00:00:51.066 --> 00:00:54.233
<c.white>and very little data for corpus-driven research exists.</c>

00:00:55.200 --> 00:00:56.700
<c.white>To address this lack of data,</c>

00:00:56.700 --> 00:01:00.866
<c.white>the DGS Corpus was created as part of a 15 year project</c>

00:01:00.866 --> 00:01:02.833
<c.white>that was started in 2009.</c>

00:01:03.299 --> 00:01:06.700
<c.white>It is an annotated reference corpus of German Sign Language,</c>

00:01:06.700 --> 00:01:10.266
<c.white>consisting of 560 hours of conversations.</c>

00:01:10.666 --> 00:01:13.466
<c.white>It includes both sign by sign annotations</c>

00:01:13.466 --> 00:01:16.233
<c.white>and translations into German and English.</c>

00:01:17.066 --> 00:01:20.766
<c.white>50 hours of that reference corpus are also publicly available</c>

00:01:20.766 --> 00:01:22.966
<c.white>as the <i>Public DGS Corpus</i>.</c>

00:01:26.066 --> 00:01:30.900
<c.white>While small when compared to many corpora of well-resourced spoken languages,</c>

00:01:30.900 --> 00:01:33.533
<c.white>the DGS Corpus and its public subset</c>

00:01:33.533 --> 00:01:37.233
<c.white>are both amongst the largest sign language corpora of their kind.</c>

00:01:37.833 --> 00:01:40.133
<c.white>One reason for this is that signed languages</c>

00:01:40.133 --> 00:01:43.099
<c.white>have no commonly used written forms</c>

00:01:43.099 --> 00:01:45.866
<c.white>and phonetic transcription is very complex.</c>

00:01:46.866 --> 00:01:51.666
<c.white>Annotating signed utterances properly is a <i>very</i> time-consuming process.</c>

00:01:52.633 --> 00:01:54.466
<c.white>One hour of recording needs about</c>

00:01:54.466 --> 00:01:58.299
<c.white>800 to 1000 hours of work before publication</c>

00:01:58.299 --> 00:02:01.833
<c.white>setting hard constraints for how much of the reference corpus</c>

00:02:01.833 --> 00:02:04.333
<c.white>could be included in the public corpus.</c>

00:02:05.700 --> 00:02:07.766
<c.white>Publishing 50 hours of data</c>

00:02:07.766 --> 00:02:11.933
<c.white>allows us to present a cross-section of the overall corpus,</c>

00:02:11.933 --> 00:02:16.099
<c.white>giving a good impression of the different kinds of contents that it covers.</c>

00:02:18.166 --> 00:02:21.133
<c.white>During the creation and publication of the corpus,</c>

00:02:21.133 --> 00:02:23.433
<c.white>the team followed established best practices</c>

00:02:23.433 --> 00:02:26.599
<c.white>for ethical research relating to deaf populations.</c>

00:02:27.599 --> 00:02:30.733
<c.white>These practices are in line with the recently introduced</c>

00:02:30.733 --> 00:02:33.233
<c.white>CARE principles of open data ethics.</c>

00:02:36.800 --> 00:02:40.300
<c.white>To see how exactly the DGS Corpus implements CARE,</c>

00:02:40.300 --> 00:02:41.933
<c.white>let’s start at the beginning.</c>

00:02:45.566 --> 00:02:48.633
<c.white>The primary stakeholders in the DGS language community</c>

00:02:48.633 --> 00:02:51.233
<c.white>are members of the deaf community in Germany.</c>

00:02:51.699 --> 00:02:55.400
<c.white>Following the principle “nothing about us without us”,</c>

00:02:55.400 --> 00:02:58.699
<c.white>the corpus project has always included deaf team members.</c>

00:02:59.400 --> 00:03:02.633
<c.white>In addition, a focus group of deaf users was formed</c>

00:03:02.633 --> 00:03:04.599
<c.white>to guide project decisions</c>

00:03:04.599 --> 00:03:07.566
<c.white>and assist us in connecting with the deaf community</c>

00:03:07.566 --> 00:03:10.233
<c.white>as well as keeping it informed about the project.</c>

00:03:12.933 --> 00:03:15.300
<c.white>To ensure <i>collective benefit</i>,</c>

00:03:15.300 --> 00:03:17.699
<c.white>the corpus was designed so that it would be</c>

00:03:17.699 --> 00:03:22.566
<c.white>both a source for linguistic research <i>and</i> a record of deaf culture,</c>

00:03:22.566 --> 00:03:24.966
<c.white>covering general life experience,</c>

00:03:24.966 --> 00:03:27.066
<c.white>deaf-specific experiences,</c>

00:03:27.066 --> 00:03:29.266
<c.white>perception of historical events,</c>

00:03:29.266 --> 00:03:31.933
<c.white>but also things like telling jokes.</c>

00:03:33.633 --> 00:03:35.866
<c.white>The goal was to create a resource</c>

00:03:35.866 --> 00:03:38.433
<c.white>that would be entertaining, informative,</c>

00:03:38.433 --> 00:03:41.633
<c.white>and that would support the identity of the community.</c>

00:03:48.199 --> 00:03:52.933
<c.white>The project recorded 330 participants from all across Germany</c>

00:03:52.933 --> 00:03:56.199
<c.white>whose primary language of daily life was DGS.</c>

00:03:57.300 --> 00:04:00.166
<c.white>Following the <i>authority to control</i> principle,</c>

00:04:00.166 --> 00:04:03.533
<c.white>informed consent was requested from all participants.</c>

00:04:04.133 --> 00:04:08.699
<c.white>This involved providing information in both DGS and German</c>

00:04:08.699 --> 00:04:10.766
<c.white>regarding the goals of the project,</c>

00:04:10.766 --> 00:04:12.166
<c.white>uses of the data,</c>

00:04:12.166 --> 00:04:14.466
<c.white>and what the rights of the participants are.</c>

00:04:15.633 --> 00:04:20.100
<c.white>These rights include restricting for what purposes the data may be shared</c>

00:04:20.100 --> 00:04:24.566
<c.white>and also reviewing the recordings to give or withhold their approval</c>

00:04:24.566 --> 00:04:27.800
<c.white>for either entire recordings or individual moments.</c>

00:04:33.433 --> 00:04:35.466
<c.white>Let’s fast forward a few years.</c>

00:04:35.899 --> 00:04:39.433
<c.white>After a lot of work annotating and translating recordings,</c>

00:04:39.433 --> 00:04:44.333
<c.white>the first full release of the public corpus was published in 2018.</c>

00:04:45.866 --> 00:04:48.899
<c.white>We also release updated versions on a regular basis</c>

00:04:48.899 --> 00:04:50.966
<c.white>to add more data, make corrections,</c>

00:04:50.966 --> 00:04:52.633
<c.white>and to react to feedback.</c>

00:04:55.300 --> 00:04:56.699
<c.white>As I mentioned before,</c>

00:04:56.699 --> 00:04:59.566
<c.white>the DGS Corpus is both a linguistic resource</c>

00:04:59.566 --> 00:05:01.399
<c.white>and a record of deaf culture.</c>

00:05:01.800 --> 00:05:04.033
<c.white>So to maximise its <i>collective benefit</i></c>

00:05:04.033 --> 00:05:07.000
<c.white>we released its data on two separate portals.</c>

00:05:14.733 --> 00:05:17.266
<c.white>The first one is <i>My DGS</i>,</c>

00:05:17.266 --> 00:05:19.333
<c.white>a community portal for deaf people</c>

00:05:19.333 --> 00:05:22.500
<c.white>and others interested in DGS and deaf culture.</c>

00:05:22.933 --> 00:05:26.399
<c.white>It provides all recordings with optional German subtitles</c>

00:05:26.399 --> 00:05:28.800
<c.white>and its design focusses on making it easy</c>

00:05:28.800 --> 00:05:31.566
<c.white>to find and watch interesting content.</c>

00:05:32.033 --> 00:05:33.600
<c.white>Here is a little example.</c>

00:05:53.666 --> 00:05:56.866
<c.white>The second portal is My DGS – annotated,</c>

00:05:56.866 --> 00:05:59.633
<c.white>a research portal that provides the same recordings</c>

00:05:59.633 --> 00:06:03.766
<c.white>with full sign annotations and translations in German and English.</c>

00:06:04.199 --> 00:06:06.399
<c.white>All data is available to download,</c>

00:06:06.399 --> 00:06:09.533
<c.white>but can also be viewed in an online transcript viewer.</c>

00:06:09.833 --> 00:06:11.500
<c.white>Here is the video from before</c>

00:06:11.500 --> 00:06:13.833
<c.white>Here is the video from before as seen through the research portal.</c>

00:06:33.566 --> 00:06:36.199
<c.white>The research portal also provides a type index</c>

00:06:36.199 --> 00:06:39.466
<c.white>of all unique signs occurring in the corpus.</c>

00:06:47.466 --> 00:06:51.333
<c.white>For each sign you receive an overview of its corpus occurrences,</c>

00:06:51.333 --> 00:06:53.100
<c.white>grouped by sign sense,</c>

00:07:02.566 --> 00:07:07.633
<c.white>and where possible a studio recording and phonetic transcription of its citation form</c>

00:07:07.633 --> 00:07:10.533
<c.white>as well as links to other lexical resources.</c>

00:07:29.366 --> 00:07:31.666
<c.white>The publication of the portals is also</c>

00:07:31.666 --> 00:07:34.800
<c.white>where FAIR joins CARE in our considerations.</c>

00:07:35.399 --> 00:07:37.733
<c.white>To make them reliably findable,</c>

00:07:37.733 --> 00:07:41.733
<c.white>each portal is treated as a separate but related dataset</c>

00:07:41.733 --> 00:07:43.800
<c.white>and given separate DOIs.</c>

00:07:44.399 --> 00:07:47.466
<c.white>For a simpler dataset, a single DOI would be sufficient,</c>

00:07:47.466 --> 00:07:51.300
<c.white>but for a complex dataset like the <i>Public DGS Corpus</i></c>

00:07:51.300 --> 00:07:56.000
<c.white>we find it advisable to also have identifiers for individual parts.</c>

00:07:59.699 --> 00:08:03.633
<c.white>So we create DOIs for each individual transcript</c>

00:08:03.633 --> 00:08:06.866
<c.white>as well as for each type in the type index.</c>

00:08:07.233 --> 00:08:10.266
<c.white>That way researchers can clearly specify</c>

00:08:10.266 --> 00:08:13.833
<c.white>which transcripts or signs they refer to in their research.</c>

00:08:15.800 --> 00:08:19.966
<c.white>On top of that, whenever we release an updated version of the corpus,</c>

00:08:19.966 --> 00:08:23.733
<c.white>every component that has changed also receives a new DOI,</c>

00:08:23.733 --> 00:08:28.100
<c.white>so it is always clear which version of the corpus is being referred to.</c>

00:08:31.300 --> 00:08:34.433
<c.white>All of these DOIs are then given qualified references</c>

00:08:34.433 --> 00:08:37.666
<c.white>to clarify how they are related to each other.</c>

00:08:45.799 --> 00:08:49.899
<c.white>Each DOI also comes with a set of machine-readable metadata</c>

00:08:49.899 --> 00:08:52.233
<c.white>covering general dataset information</c>

00:08:52.233 --> 00:08:54.933
<c.white>like its name, authors, release date,</c>

00:08:54.933 --> 00:08:57.733
<c.white>and all those qualified references I just mentioned.</c>

00:09:02.566 --> 00:09:05.633
<c.white>For information that is more specific to language data</c>

00:09:05.633 --> 00:09:08.633
<c.white>we also provide a CMDI file for each transcript.</c>

00:09:08.933 --> 00:09:12.366
<c.white>In there we specify metadata about the participants,</c>

00:09:12.366 --> 00:09:13.966
<c.white>elicitation tasks,</c>

00:09:13.966 --> 00:09:16.633
<c.white>the languages of the primary and secondary data,</c>

00:09:16.633 --> 00:09:17.966
<c.white>and so on.</c>

00:09:23.766 --> 00:09:27.666
<c.white>Apart from this metadata, the corpus has a wealth of documentation.</c>

00:09:27.866 --> 00:09:29.733
<c.white>In addition to peer-reviewed articles,</c>

00:09:29.733 --> 00:09:34.399
<c.white>there are over thirty project notes documenting various aspects of the project.</c>

00:09:42.799 --> 00:09:44.600
<c.white>This includes a Data Statement,</c>

00:09:44.600 --> 00:09:47.866
<c.white>a document type specifically designed to help researchers</c>

00:09:47.866 --> 00:09:50.333
<c.white>understand the background of a dataset</c>

00:09:50.333 --> 00:09:52.666
<c.white>and anticipate its inherent biases.</c>

00:09:53.233 --> 00:09:57.366
<c.white>Of course, all project notes have their own version-controlled DOIs.</c>

00:10:03.899 --> 00:10:05.566
<c.white>Another challenge for the project</c>

00:10:05.566 --> 00:10:08.933
<c.white>was the long term archival of its original recordings.</c>

00:10:09.899 --> 00:10:13.033
<c.white>At over one thousand hours of raw recording time,</c>

00:10:13.033 --> 00:10:14.700
<c.white>captured from several perspectives</c>

00:10:14.700 --> 00:10:18.566
<c.white>by 7 different high definition or stereoscopic cameras,</c>

00:10:18.566 --> 00:10:24.333
<c.white>the original data collection resulted in 730,000 gigabytes of video files.</c>

00:10:27.100 --> 00:10:30.100
<c.white>An additional 250,000 gigabytes of data</c>

00:10:30.100 --> 00:10:33.466
<c.white>have so far been produced through secondary data collections,</c>

00:10:33.466 --> 00:10:36.700
<c.white>studio recordings made for the presentation of the corpus,</c>

00:10:36.700 --> 00:10:39.433
<c.white>and recordings made for a corpus-based dictionary</c>

00:10:39.433 --> 00:10:41.466
<c.white>that is also part of the project.</c>

00:10:43.933 --> 00:10:46.766
<c.white>That leaves us with a whole petabyte of data</c>

00:10:46.766 --> 00:10:49.566
<c.white>that needs to be securely backed up and archived,</c>

00:10:49.566 --> 00:10:51.933
<c.white>ideally following FAIR principles.</c>

00:10:53.399 --> 00:10:57.066
<c.white>Luckily, the Research Data Repository of Hamburg University</c>

00:10:57.066 --> 00:11:00.233
<c.white>provides a decentralised redundant server structure</c>

00:11:00.233 --> 00:11:03.366
<c.white>capable of handling large data collections like ours.</c>

00:11:04.899 --> 00:11:08.200
<c.white>The archive is run using Invenio RDM,</c>

00:11:08.200 --> 00:11:10.799
<c.white>the same repository software as Zenodo,</c>

00:11:10.799 --> 00:11:13.933
<c.white>which enables persistent storage, fast access,</c>

00:11:13.933 --> 00:11:16.899
<c.white>associated metadata, and clear versioning.</c>

00:11:18.833 --> 00:11:22.533
<c.white>While the original recordings themselves are not publicly accessible,</c>

00:11:22.533 --> 00:11:24.600
<c.white>their identity and metadata are,</c>

00:11:24.600 --> 00:11:27.533
<c.white>allowing us to also link each of them explicitly</c>

00:11:27.533 --> 00:11:31.399
<c.white>with the transcripts of the public corpus that originate from them.</c>

00:11:37.533 --> 00:11:40.233
<c.white>And that brings me to the end of my presentation.</c>

00:11:40.600 --> 00:11:42.600
<c.white>Of course, I only skimmed the surface</c>

00:11:42.600 --> 00:11:44.033
<c.white>and even skipped a few aspects,</c>

00:11:44.033 --> 00:11:47.633
<c.white>such as usage licences and anonymisation,</c>

00:11:47.633 --> 00:11:51.600
<c.white>so if you have any questions now or later, I’d be happy to answer them.</c>

00:11:51.933 --> 00:11:55.033
<c.white>And of course please go and check out the corpus itself</c>

00:11:55.033 --> 00:11:57.000
<c.white>and the many stories it contains.</c>

00:11:57.466 --> 00:11:58.366
<c.white>Thank you!</c>