{ "paper_id": "2020", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T12:26:50.070347Z" }, "title": "Addressing Cha(lle)nges in Long-Term Archiving of Large Corpora", "authors": [ { "first": "Denis", "middle": [], "last": "Arnold", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Bernhard", "middle": [], "last": "Fisseni", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Pawe\u0142", "middle": [], "last": "Kamocki", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Oliver", "middle": [], "last": "Schonefeld", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Marc", "middle": [], "last": "Kupietz", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Thomas", "middle": [], "last": "Schmidt", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper addresses long-term archival for large corpora. Three aspects specific to language resources are focused, namely (1) the removal of resources for legal reasons, (2) versioning of (unchanged) objects in constantly growing resources, especially where objects can be part of multiple releases but also part of different collections, and (3) the conversion of data to new formats for digital preservation. It is motivated why language resources may have to be changed, and why formats may need to be converted. As a solution, the use of an intermediate proxy object called a signpost is suggested. The approach will be exemplified with respect to the corpora of the Leibniz Institute for the German Language in Mannheim, namely the German Reference Corpus (DeReKo) and the Archive for Spoken German (AGD).", "pdf_parse": { "paper_id": "2020", "_pdf_hash": "", "abstract": [ { "text": "This paper addresses long-term archival for large corpora. Three aspects specific to language resources are focused, namely (1) the removal of resources for legal reasons, (2) versioning of (unchanged) objects in constantly growing resources, especially where objects can be part of multiple releases but also part of different collections, and (3) the conversion of data to new formats for digital preservation. It is motivated why language resources may have to be changed, and why formats may need to be converted. As a solution, the use of an intermediate proxy object called a signpost is suggested. The approach will be exemplified with respect to the corpora of the Leibniz Institute for the German Language in Mannheim, namely the German Reference Corpus (DeReKo) and the Archive for Spoken German (AGD).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The current paper investigates long-term archival (LTA) for large corpora, specifically corpora that are constantly extended and contain material where the conglomerate of commercial interests, intellectual property rights and privacy rights constitutes a non-trivial problem; we call them growing corpora. We focus on three aspects of archiving growing corpora which are related to changing resources, and in our opinion, can be approached by using tombstones or, preferably, signposts. While we restrict our attention to growing corpora, all aspects apply to other kinds of corpora as well, but generally to a different degree. In the interest of space, we leave it to the attentive reader to judge the applicability. Caron et al. (2017) discuss their solution in the context of the Open Archival Information System model (generally abbreviated: OAIS, see CCSDS, 2012; for an overview, see also the 4th chapter by O\u00dfwald in Neuroth et al., 2009) , and specifically the aspect of dealing with the ingest of a submission information package into the archive. In the OAIS, an edition is characterized by change in the content (e.g., O\u00dfwald only speaks of additions), while a version is the result of a migration (cf. CCSDS, 2012, p. 1-9; \u00a7 5.1, esp. \u00a75.1.3.4 Transformation). In this sense, while initially motivated by transparently dealing with editions, the current approach tries to integrate versions and editions into a common model. Reproducibility in science in general and reuse of data in particular have been recognised as important goals over the last years, also connected to the adoption of the term and concept of open science (cf. for an overview Fecher and Friesike, 2014) and the publication and wide-spread adoption of the FAIR principles (Wilkinson et al., 2016) . To maintain reproducibility of a documented scientific procedure, it also is necessary to maintain the access to some form of the data. This is in immediate conflict with the removal of data. Surprisingly, while the change of data regarding formats is generally considered in the context of preservation (cf. an example for research data Conway et al., 2011) , we have found no example that considers the change of data sets due to legal issues. One aspect that distinguishes growing corpora from others is that for legal reasons, it may be necessary to remove or modify part of the data. In long-term archival (see, e.g., Digital Preservation Coalition, 2015; Neuroth et al., 2009) , this case is generally neglected, as it is normally assumed that data will stay unchanged. To remain as close to the spirit of LTA as possible, one will still want to deliver some useful information when someone tries to retrieve the removed objects. This is especially important since it is reasonable to assume that a researcher will not expect that data referenced with some form of persistent identifier has been altered. To our knowledge, only Caron et al. (2017) have considered the deletion of objects in LTA. They describe a 'tombstone' which steps in for objects that needed to be deleted for legal reasons. Systems like Fedora Commons or DSpace allow for removal of resources and provide tombstone objects. 1 We will discuss the differences between these and our approaches below. Another aspect is parsimonious representation of data in corpora with many releases: Objects may be referenced in different releases and resources. Corpora that are curated in projects where the corpora are constantly extended and published in frequent releases will have many (unchanged) objects in common across different releases. Furthermore, an object may belong to different collections. Generally, a digital long-term archive will avoid storing the same digital objects multiple times. Keeping only one copy per object ensures that there is no confusion about the state of an object and storage space is not wasted, especially when objects are considerably large. The third and last aspect is specific to long-time preservation: It is unforeseeable if and when a given file format may become deprecated. But once this is the case, the archive will have to migrate the respective files to the new format and make them accessible along with the original files. This can be seen as a departure from the original model, which states \"[\u2026] that the new archival implementation of the information is a replacement for the old\" (CCSDS, 2012, p. 1-11) . To explain the proposal, we distinguish the notions of conceptual object (CO) and logical object (LO) (see chapter 9 by Stefan Funk in Neuroth et al., 2009) . 2 A CO can be realized in different LOs, for instance an audio recording (CO) can be realized in files of different audio formats (LO). The first two cases, i.e. removal and and versioning primarily concern changes of conceptual objects -although the changes will be mirrored in LOs -, while the case of format conversion only concerns logical objects. This observation will form the basis of our technical proposal.", "cite_spans": [ { "start": 720, "end": 739, "text": "Caron et al. (2017)", "ref_id": "BIBREF5" }, { "start": 926, "end": 947, "text": "Neuroth et al., 2009)", "ref_id": null }, { "start": 1662, "end": 1688, "text": "Fecher and Friesike, 2014)", "ref_id": "BIBREF11" }, { "start": 1757, "end": 1781, "text": "(Wilkinson et al., 2016)", "ref_id": "BIBREF25" }, { "start": 2122, "end": 2142, "text": "Conway et al., 2011)", "ref_id": "BIBREF8" }, { "start": 2395, "end": 2444, "text": "(see, e.g., Digital Preservation Coalition, 2015;", "ref_id": null }, { "start": 2445, "end": 2466, "text": "Neuroth et al., 2009)", "ref_id": null }, { "start": 2918, "end": 2937, "text": "Caron et al. (2017)", "ref_id": "BIBREF5" }, { "start": 3186, "end": 3187, "text": "1", "ref_id": null }, { "start": 4387, "end": 4409, "text": "(CCSDS, 2012, p. 1-11)", "ref_id": null }, { "start": 4547, "end": 4568, "text": "Neuroth et al., 2009)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction: Three Challenges", "sec_num": "1." }, { "text": "The Leibniz Institute for the German Language (IDS) is building up a long-term archiving repository for linguistic data. Current work is focusing on the development of appropriate ways of ingesting the IDS's own corpora of written and spoken language. Both can be viewed as exemplars of large and growing corpora: The German Reference Corpus DeReKo (Kupietz et al., 2010 (Kupietz et al., , 2018 has been built at IDS since its foundation in the mid-60s (Teubert and Belica, 2014) . It currently contains 46.9 billion tokens (Leibniz-Institut f\u00fcr Deutsche Sprache, 2020) corresponding to 56 GB disk space (without automatic annotations) and is used by more than 40 000 German linguists world-wide, primarily via specialized analysis platforms, such as COSMAS II (Bodmer, 2005) and KorAP (Ba\u0144ski et al., 2013) . The Archive for Spoken German (AGD; Schmidt, 2017) hosts around 80 corpora of spoken language. The digital available corpora are published via the Datenbank f\u00fcr Gesprochenes Deutsch (Schmidt and Gasch, 2019) . While much smaller than their written counterparts in terms of number of documents or tokens of (transcribed) text, they can also be viewed as large corpora because they comprise large digital audio or video files. For example, the 1 113 recordings of the BOLSA study (Lehr and Thomae, 1987 ) make up for altogether 2 833 hours of audio. Stored as mono WAV files with a sampling rate of 48 kHz, they occupy around 1 TB of disk space. For the latest version of the FOLK corpus (Schmidt, 2016) , the textual data amounts to around 2.5 million transcribed tokens (less than 0.5 GB), whereas the archived media data (stereo WAV, 48 kHz for audio, MPEG-4 in a resolution of 1 980 \u00d7 1 080 for video) is also around 1 TB.", "cite_spans": [ { "start": 349, "end": 370, "text": "(Kupietz et al., 2010", "ref_id": "BIBREF13" }, { "start": 371, "end": 394, "text": "(Kupietz et al., , 2018", "ref_id": "BIBREF15" }, { "start": 453, "end": 479, "text": "(Teubert and Belica, 2014)", "ref_id": "BIBREF23" }, { "start": 761, "end": 775, "text": "(Bodmer, 2005)", "ref_id": "BIBREF1" }, { "start": 786, "end": 807, "text": "(Ba\u0144ski et al., 2013)", "ref_id": "BIBREF0" }, { "start": 846, "end": 860, "text": "Schmidt, 2017)", "ref_id": "BIBREF21" }, { "start": 992, "end": 1017, "text": "(Schmidt and Gasch, 2019)", "ref_id": "BIBREF27" }, { "start": 1288, "end": 1310, "text": "(Lehr and Thomae, 1987", "ref_id": "BIBREF16" }, { "start": 1496, "end": 1511, "text": "(Schmidt, 2016)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Background", "sec_num": "1.1." }, { "text": "There are several possible scenarios where parts of large corpora intended for long-term archiving have to be deleted for legal reasons. Three legal frameworks seem to be of particular relevance here: intellectual property (IP), data protection and criminal law.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Legal Aspects", "sec_num": "1.2." }, { "text": "Firstly, concerning IP, it is important to note that language data are, for the most part, protected by copyright. As such, their use (i.e. reproduction and communication to the public) is lawful only in one of two cases: (1) a permission (license) has been obtained from the right holder or (2) the use is covered by a statutory exception or limitation (e.g. for teaching and research). In both cases, long-term archiving may be impacted. A license can be granted for limited duration only, and once it comes to its term, the work can no longer be lawfully used.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Intellectual Property", "sec_num": null }, { "text": "Technically speaking, it does not have to be deleted, but any further copying (even in a computer's memory) would amount to copyright infringement. Although from the user's point of view it is advantageous to use licenses that are not limited in time (or, rather, are granted for the whole duration of copyright), such as Creative Commons licenses, right holders cannot be forced to grant them. In practice, it is usually the case that the longer the licence period, the less likely it is that the licence will be granted, or higher fees may be charged, or both (see Kupietz et al., 2014 , for a more detailed discussion of the trade-offs). In the case of DeReKo, the typical duration of commercial licenses is one year, while donated licenses are almost always unlimited in time. A license can also be revoked by the licensor -usually because the right of revocation is specifically provided for in the license itself, in which case, of course, the stipulated modalities (e.g. prior notice) would have to be respected. Creative Commons licenses, for example, terminate automatically upon any breach of the license's terms. It is also possible, albeit in very limited cases, that the statute grants the right holder the right of revocation (i.e. the license can be revoked even if it does not stipulate so). Under German law, for example, an exclusive license can be revoked if the work is not used by the licensee ( \u00a741 UrhG 3 ). More interestingly, still under German law, the author has the right to revoke a license 'for changed conviction' ( \u00a742 UrhG), i.e. when he decides that the work no longer reflects his conviction. In this case, the author has to adequately compensate the licensee for the revocation, and if in the future he decides to use the work again, he shall grant a new license to the licensee 'on reasonable conditions'. Exceptionally, the right of revocation for changed conviction can also be exercised by the author's heirs; then, however, it would require a proof that the author would have exercised exercised the right prior to his death. Upon revocation, whether on contractual or statutory grounds, the licensed data can no longer be lawfully used. In the case of DeReKo, many licenses are explicitly revocable at any time with a period of a few weeks. Since the 2000s at the latest, a corresponding addition to the license conditions has often proved necessary in order to be able to conclude license agreements at all and in a reasonable time. So far, however, no licensor has made use of his right of revocation. Statutory exceptions may seem to provide for a more stable ground for long-term archiving, but it is not always the case. It should be kept in mind that the exception may simply not allow for long-term archiving (such as the current data mining exception in German law - \u00a760d UrhG), and even if it does, it may simply cease to apply at some point, or be replaced by a different, stricter norm (even though in the past decade or two the trend seems to be towards broadening the scope of statutory exceptions). Moreover, albeit very rarely, an exception may come with an 'expiry date' -this is the case of exceptions introduced in German law by the UrhWissG 4 (covering such uses as teaching, research, data mining and uses made by libraries), which will cease to apply at the end of February 2023 (although they are expected to be either maintained by the legislator, or replaced by other similar exceptions. Therefore, when long-term archiving is based on statutory exceptions, it is of the essence to stay informed about the developments in the legal framework, and adjust the archiving policy accordingly.", "cite_spans": [ { "start": 567, "end": 587, "text": "Kupietz et al., 2014", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Intellectual Property", "sec_num": null }, { "text": "Another legal framework that crucially impacts long-term archiving is data protection; the most important source of data protection law is the famous General Data Protection Regulation (GDPR) which entered into application in the European Union on 25 May 2018. First of all, it should be mentioned that when the corpus contains personal data (which is not unlikely to happen, taking into account the large scope of the notion), archiving for an undefined (and potentially unlimited) duration may simply not be an option. Storage limitation (the principle according to which personal data can be kept for longer than necessary to achieve the purposes of the processing -Article 5(1)(e) of the GDPR) is one of the fundamental principles governing the processing of personal data under the GDPR. However, the GDPR allows for derogations from this principle when the processing is carried out solely 'for archiving purposes in the public interest', or for research or statistical purposes (Article 89 of the GDPR). To be able to qualify for the derogation, however, the processing has to be subject to 'appropriate safeguards' such as e.g. pseudonymization. The rules regarding these purposes of processing remain largely country-specific -in Germany, at the federal level, processing for research purposes is governed by \u00a727 of the BDSG 5 , archiving in the public interest by \u00a728, and 'appropriate safeguards' are listed in \u00a722 of the same Act. Even if the storage limitation principle with its derogations is observed, some data may still have to be deleted on the grounds of data protection. When the processing is based on consent of the data subject (which, alongside 'legitimate interest' is probably the most common ground for the processing of personal data in language corpora), the consent can be withdrawn at any time (Article 7(3) of the GDPR). The withdrawal has no retroactive effect (i.e. the processing based on consent prior to its withdrawal does not 'become unlawful'), but any further processing should stop (although it is possible to resume processing on a different ground, e.g. based on legitimate interest). However, if the data is processed on the ground of legitimate interest, the data subject may still exercise the right to object (Article 21 of the GDPR), in which case the processing should stop, unless the controller (the person or entity who defines the means and purposes of the processing) demonstrates 'compelling legitimate grounds' for the processing which override the interests, rights and freedoms of the data subject. The right to object does not apply to processing for research purposes, or for the purpose of 'archiving in the public interest' (Article 89 of the GDPR, \u00a7 \u00a7 27-28 of the BDSG). Finally, the data subject may also exercise his right of erasure (Article 17 of the GDPR), commonly referred to as 'the right to be forgotten'. This right is not limited even when the processing is carried out for research or archiving purposes (unless it 'seriously impairs' these purposes); however, perhaps contrary to the common belief, the conditions for exercising this right are in fact very strict, and in practice seem to require some sort of prior violation of the GDPR on behalf of the controller. Perhaps most importantly, the data subject may request erasure if the data minimisation principle has been violated, i.e. the data are no longer necessary to achieve the purpose of the processing. This ground can be successfully used while requesting erasure e.g. from search engines, and possibly also from large, publicly accessible corpora. It is worth noting that if the request for erasure is well-founded, the controller shall also make reasonable steps (including technical measures) to inform other controllers who process the same information, so that they can also proceed with the erasure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Protection", "sec_num": null }, { "text": "Last but not least, criminal law, and more specifically the rules regarding defamation and defamation-related offences (slander, libel, insult\u2026), may also require deletion of some parts of large corpora. This is especially relevant for newspapers or other press materials. Since national rules may vary significantly (there is no harmonised law of defamation at the EU level), we will use \u00a7 186 of the German Criminal Code for illustration purposes. The text provides that whoever disseminates a fact related to another person which may defame him or negatively affect public opinion about him, is punished with a fine or imprisonment for up to one year (when the offence is committed by dissemination of written materials, the penalty increases to two years). Apart from that, the claimant may also obtain an injunction (i.e. a court order for the defendant to stop disseminating the material), even preliminary (i.e. applicable even before the final decision on the merits of the case is made by the court). Needless to say, a defamation claim, even ill-founded may lead to (at least temporary) deletion of parts of long-term archived corpora. In the case of DeReKo, injunctions are by far the most common reason for the removal of individual texts, with about two incidents per week. The obligation to remove the texts is also stipulated in the license agreements with the right holders. There is a consensus within the German linguistic community that the removal of individual texts is unavoidable and typically irrelevant with respect to linguistic findings and should not pose an insoluble problem in terms of reproducibility of research results, which is reflected in the guidelines on legal aspects of handling corpora of the German Research Foundation DFG (Deutsche Forschungsgemeinschaft, 2015, p. 19) , quoted here from its English translation:", "cite_spans": [ { "start": 1776, "end": 1812, "text": "Forschungsgemeinschaft, 2015, p. 19)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Criminal Law", "sec_num": null }, { "text": "Regarding the still existing problem of persistence of research data, there is a certain pragmatic consensus within the scientific community: text deletions because of personality rights should be considered acceptable also epistemologically, since the replicability of important and methodically valid research results does not depend on individual texts. What is probably more important is de facto the organizational effort that can be caused by individual deletions. It is recommended to factor this into project costs in advance, if possible. (Wildgans et al., 2017, p. 20) The three frameworks presented above may to a limited extent be derogated from by laws on public archives, such as the Bundesarchivgesetz or Landesarchivgesetze in Germany (socalled L\u00f6schungssurrogat), or Code du patrimoine in France. However, apart from them being heavily country-specific, their application is usually limited to 'official' registries or other documents of key importance for public administration, or to archiving by specifically designated institutions. Therefore, it is our opinion that the relevance of such laws on public archives for long-term archiving of language corpora is very limited and so they fall outside of the scope of this paper.", "cite_spans": [ { "start": 548, "end": 578, "text": "(Wildgans et al., 2017, p. 20)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Criminal Law", "sec_num": null }, { "text": "There are two cases in which COs change, only one of which constitutes a veritable challenge. The other one can be seen as unproblematic. The unproblematic case arises when a resource is published in a new version, e.g. containing more annotations, but also correcting mistakes that will stay accessible in the previous version. The long-term archive will in this case simply issue a new version of the CO and the old version stays intact and accessible. There is a certain conflict of interests here: On the one hand, it may be interesting to users to see that there is a new version of an CO; on the other hand, integrating this information into the archive would most evidently be possible changing the metadata, a measure which evidently goes against the general guarantees of long-term archival. We suggest that in this case the latter point far outweighs the former: It is not necessary to point to the new version in the long-term archive and make changes to metadata. However, it is by perfectly admissible from an LTA perspective to point to old versions from the new ones -as long as the latter are archived after the former have been, as is normally the case. To improve usability, the presentation layer of the archive can invert these links without integrating them into the metadata proper. The interesting case is when parts of a corpus must be altered; this generally occurs due to legal reasons: An alternative version of the respective CO and its LOs must be created, or the CO must even be deleted completely. In any case, the old version will no longer be accessible. Caron et al.'s focus are single packages of digitised documents (images, text files) rather than growing, hierarchical corpora and partial modifications. Based on the OAIS terminology explained in our first section, they determine whether there is a new version or a new edition as follows (see their figure 3): Disregarding ingest failures, a modification or deletion of data (in their case, e.g., improved imaging) leads to a new version, while additional content or modification of metadata leads to a new edition.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Versioning of COs", "sec_num": "1.3." }, { "text": "Finally, it may happen that a certain file format falls out of use. For instance, in the area of video formats, Apple has retired its QuickTime format in macOS 10.15 (Catalina). In the area of text annotation, SGML (ISO8879:1986 , 1986 has given way to XML (Bray et al., 1997) . To anecdotically trace one of our migration paths: DeReKo used SGML/CES between 1999 and 2005, and was consequently converted to an XMLbased format (for the history and the decisions involved, see L\u00fcngen and Sperberg-McQueen, 2012) , first based on the TEI's P3 recommendations (Sperberg-McQueen and Burnard, 1999) , later converted to TEI P5 (Burnard and Bauman, 2020) , the customization of the annotation is called I5. Similarly, in the area of spoken language annotation, the IDS is in the process of switching to the new ISO standard ISO-24624:2016 (ISO, 2016) for new annotations, s, and has also converted old annotations from a variety of formats including SGML, HTML and plain text. Also, the tools developed at the IDS, especially the EXMARaLDA family (Schmidt and W\u00f6rner, 2014) ) are being adapted to work with this format. One can argue that the older SGML and XML formats are still usable, but the modern formats provide much better interoperability at the current time. So while formats used in longterm archival are generally selected to minimize the chance of complete obsolescence or loss of readability, it may still be preferable to provide additional formats that are more readily supported by contemporary software. In this case, the original LO is not completely replaced, but it may be preferable to deliver an object in another format if one queries for the CO. Again it is inconvenient and misleading to modify the metadata of parent resources, as conceptually, not the CO but only the files realizing it have changed. An important question is reversibility of transformations. For formats like video, where we store lossily compressed data, a transformation would probably not be reversible (CCSDS, 2012, p. 5-6f), as transcoding would introduce new compression COs. In case of the conversion of DeReKo from SGML to XML (cf. L\u00fcngen and Sperberg-McQueen, 2012) , however, the migration was reversible in the sense that all old data could be losslessly translated back. I5 has evolved further (as of this writing, the latest version is from 2020-03-05) 6 and accommodates features for new kinds of text (such as computer-mediated communication), new data may not be retranslatable to the old format. This effect of the migration path illustrates a further complication regarding growing corpora. With respect to the OAIS model, we can model this as three cases: (1) First, the conversion is merely a change of the Dissemination Information Package (DIP), or (2) alternatively, it may constitute a migration, namely a a transformation (CCSDS, 2012, \u00a75.1.3.4). By keeping the original data, we partly transcend the OAIS model.", "cite_spans": [ { "start": 210, "end": 228, "text": "SGML (ISO8879:1986", "ref_id": null }, { "start": 229, "end": 235, "text": ", 1986", "ref_id": "BIBREF12" }, { "start": 257, "end": 276, "text": "(Bray et al., 1997)", "ref_id": "BIBREF2" }, { "start": 476, "end": 510, "text": "L\u00fcngen and Sperberg-McQueen, 2012)", "ref_id": "BIBREF17" }, { "start": 579, "end": 593, "text": "Burnard, 1999)", "ref_id": null }, { "start": 622, "end": 648, "text": "(Burnard and Bauman, 2020)", "ref_id": "BIBREF4" }, { "start": 2130, "end": 2164, "text": "L\u00fcngen and Sperberg-McQueen, 2012)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Pointing to the Converted", "sec_num": "1.4." }, { "text": "Data in a Transparent Way", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Signposts: Dealing with Modified or Deleted", "sec_num": "2." }, { "text": "In DeReKo, the structure of the corpus has three levels: corpus (i.e., subcorpus), document and ultimately text. What the corpus and document levels correspond to, depends on the text type. For newspapers, for example, a year volume corresponds to a 'corpus' and a month to a 'document'. It is in newspaper and magazine documents that a removal may occur due to injunctions for privacy reasons. We have not yet had a case where a whole volume had to be removed. For reasons of work effort, we have to retract the whole corpus release archive in case an injunction occurs. The next release archive of DeReKo, however, will contain a modified version of the document: We generally remove the body of a text, marking it as a gap in the XML annotation. (For technical reasons, it is necessary to have a