WEBVTT 00:00:00.000 --> 00:00:09.000 Thanks a lot. 00:00:09.000 --> 00:00:11.160 Thanks for the invitation and the occasion to be here. 00:00:11.160 --> 00:00:19.800 It's really wonderful to witness the start of digital humanities in earnest at FAU. 00:00:19.800 --> 00:00:20.800 Very happy to be here. 00:00:20.800 --> 00:00:29.160 And, yes, I will talk about linked open data for literary history and how linked open data 00:00:29.160 --> 00:00:35.280 may help us think differently about literary history, but also about how wanting to do 00:00:35.280 --> 00:00:40.000 literary history differently helps us think about linked open data as well. 00:00:40.000 --> 00:00:46.200 You can follow along with the slides if you'd like using the link shown here and at the 00:00:46.200 --> 00:00:49.640 bottom of the slide. 00:00:49.640 --> 00:00:51.560 So this is what I would like to cover. 00:00:51.560 --> 00:00:55.200 First of all, I will talk a little bit about text and data mining or machine learning, 00:00:55.200 --> 00:00:59.240 linked open data and literary history as like separate bits and pieces. 00:00:59.240 --> 00:01:04.880 And what I then will try to show is how we brought these three pieces together in a project 00:01:04.880 --> 00:01:08.720 that ran until last year, mining and modeling text. 00:01:08.720 --> 00:01:12.920 And mining and modeling, you already have text and data mining and linked open data 00:01:12.920 --> 00:01:15.200 modeling in there. 00:01:15.200 --> 00:01:20.040 I will just show one example of what kinds of questions you can address with this kind 00:01:20.520 --> 00:01:26.920 But there's a huge variety of data in the resulting database. 00:01:26.920 --> 00:01:29.600 So let's get right into it. 00:01:29.600 --> 00:01:35.160 Text and data mining plus LOD plus literary history equals what? 00:01:35.160 --> 00:01:37.400 But first of all, a little bit of background. 00:01:37.400 --> 00:01:42.360 I think that you could say that there are three modes of data and digital humanities. 00:01:42.360 --> 00:01:49.480 Qualitative digital humanities, where data sets are typically small, flawlessly curated, 00:01:49.560 --> 00:01:54.840 very specific domains, often heavily annotated. 00:01:54.840 --> 00:02:00.040 And this could be called smart data, because there are many intelligent, nuanced ways that 00:02:00.040 --> 00:02:01.640 we can interact with them. 00:02:01.640 --> 00:02:06.040 And the prototype of this kind of data is digital scholarly editions, especially genetic 00:02:06.040 --> 00:02:11.520 editions or critical editions, where every comma is marked up in every different variant. 00:02:11.520 --> 00:02:16.560 And so the prototypical example for me is the Faust edition that really pushed us to 00:02:16.560 --> 00:02:19.200 the limits, I think. 00:02:19.240 --> 00:02:24.240 And then there's quantitative digital humanities, where data sets are typically large, more 00:02:24.240 --> 00:02:28.880 or less annotated or only simply annotated. 00:02:28.880 --> 00:02:33.600 They may have errors, they may have all kinds of biases, but they're big and generic, so 00:02:33.600 --> 00:02:35.120 they have their uses. 00:02:35.120 --> 00:02:38.680 And a typical way of doing this kind of digital humanities would be, for example, doing a 00:02:38.680 --> 00:02:45.720 topic modeling on Project Gutenberg, the whole thing, with all its strange biases and lacking 00:02:45.720 --> 00:02:48.440 metadata and you know. 00:02:48.440 --> 00:02:52.480 But I think there's a third way for digital humanities, which is bigger, smarter data 00:02:52.480 --> 00:02:53.680 in the humanities. 00:02:53.680 --> 00:02:57.040 And I don't mean this as a compromise, you know, a little bit bigger, but not quite as 00:02:57.040 --> 00:02:58.040 well annotated. 00:02:58.040 --> 00:03:03.320 I really think we can bring together scale and nuance or detail. 00:03:03.320 --> 00:03:07.880 And one approach to do that is to use text mining, machine learning for annotation, for 00:03:07.880 --> 00:03:12.240 information extraction, and to model everything in linked open data, so that you really have 00:03:12.240 --> 00:03:18.400 contextualized knowledge and richly contextualized analysis when you perform your analysis. 00:03:19.400 --> 00:03:25.080 And one attempt to do this is mining and modeling text. 00:03:25.080 --> 00:03:26.520 So a little bit background. 00:03:26.520 --> 00:03:27.520 What is text and data mining? 00:03:27.520 --> 00:03:29.200 What is machine learning? 00:03:29.200 --> 00:03:34.320 Basically what we're trying to do with machine learning is to discover relations between 00:03:34.320 --> 00:03:39.080 surface features that we can easily recognize and more interesting features that are not 00:03:39.080 --> 00:03:44.000 so easily recognizable, but that we as humans understand, but that the machine needs to 00:03:44.000 --> 00:03:45.840 sort of learn to understand. 00:03:45.840 --> 00:03:52.120 And so we label, manually label training data that has these more interesting features, 00:03:52.120 --> 00:03:57.240 let's say direct speech or indirect speech or locations, narrative locations. 00:03:57.240 --> 00:04:04.040 And we teach the machine learning algorithm to recognize the relationship between the 00:04:04.040 --> 00:04:08.920 surface features that the machine can see and the deeper, more interesting features 00:04:08.920 --> 00:04:11.280 that we are interested in. 00:04:11.280 --> 00:04:15.040 And once the machine has learned that relationship, we can use it to annotate larger amounts of 00:04:15.040 --> 00:04:17.080 data than what we could usually do. 00:04:17.080 --> 00:04:22.320 So it's a way of getting richer information about text than the computer can do in a first 00:04:22.320 --> 00:04:23.600 pass, let's say. 00:04:23.600 --> 00:04:29.520 And this can obviously be refined bit by bit to lead to more interesting labels. 00:04:29.520 --> 00:04:33.680 And linguistic annotation is what spaCy does, it's this paradigm. 00:04:33.680 --> 00:04:39.880 But the idea is to scale it up to more literary features as well. 00:04:39.880 --> 00:04:42.860 Second part of the equation, what is literary history? 00:04:42.860 --> 00:04:46.620 And it's sort of obvious in a way, but at the same time, it's worth stating it. 00:04:46.620 --> 00:04:52.660 So the goal of literary history is to collect and document knowledge about literature over 00:04:52.660 --> 00:04:59.740 time to provide explanations of how literary systems change over time. 00:04:59.740 --> 00:05:05.100 And sometimes the biological metaphor of evolution is used for this and has its pros and cons 00:05:05.100 --> 00:05:06.180 maybe. 00:05:06.180 --> 00:05:11.900 But also to contextualize literature, like we heard from Anastasia, that there could be 00:05:12.620 --> 00:05:17.140 the conditions of production in a way, if you want, could be studied, or the reception 00:05:17.140 --> 00:05:21.700 of literature could be studied as well as part of all that literary history can be. 00:05:21.700 --> 00:05:25.300 And then there are some organizational principle of literary history, like literary history 00:05:25.300 --> 00:05:29.100 always is often organized in national literatures. 00:05:29.100 --> 00:05:33.180 Of course, also there are histories of world literature as well that transcend this, but 00:05:33.180 --> 00:05:35.300 it's often done that way. 00:05:35.300 --> 00:05:41.460 Period, movements, currents, genres, these are like big organizational units. 00:05:41.460 --> 00:05:45.780 And then of course authors and works with their themes, with the forms, with the characters, 00:05:45.780 --> 00:05:47.460 the plot, etc. 00:05:47.460 --> 00:05:52.180 And then it's typical to look at similarities and differences, so compare things and to 00:05:52.180 --> 00:05:53.820 look at how things change over time. 00:05:53.820 --> 00:05:57.500 So basically that's the scope of what we want to try to do. 00:05:57.500 --> 00:06:02.420 And it's a bit caricatural, and the field has developed and moved on beyond these basics, 00:06:02.420 --> 00:06:05.420 but that's a good starting point, I guess. 00:06:05.420 --> 00:06:07.900 And then the third element is linked open data. 00:06:07.900 --> 00:06:11.300 And this is just a very simple illustration of how this can look like. 00:06:11.300 --> 00:06:15.900 So this is a page from Wikidata. 00:06:15.900 --> 00:06:24.300 And you can see that there's Han Kang, the South Korean Nobel Prize winner of this year, 00:06:24.300 --> 00:06:28.940 who is the subject of this entry, and so also the subject of all the statements here. 00:06:28.940 --> 00:06:32.580 There's an identifier, a numerical identifier, so the name is just a label and it could be 00:06:32.580 --> 00:06:33.580 in different languages. 00:06:33.580 --> 00:06:37.140 There's a little description, and then there are the statements. 00:06:37.140 --> 00:06:40.020 And they consist of the subject, it's always the same. 00:06:40.020 --> 00:06:45.180 And then predicate, like a verb, or in this case instance of, and then the object, human, 00:06:45.180 --> 00:06:46.340 like the value. 00:06:46.340 --> 00:06:51.540 So all of linked open data is basically constructed out of these little triples, we call them, 00:06:51.540 --> 00:06:53.940 because they have three elements. 00:06:53.940 --> 00:06:58.900 And linked open data is basically a huge collection of these very, very simple statements. 00:06:58.900 --> 00:07:01.060 And I like to compare this to chess. 00:07:01.060 --> 00:07:06.900 In chess you have just a few kinds of figures and some very simple rules and a very limited 00:07:06.900 --> 00:07:08.760 field. 00:07:08.760 --> 00:07:13.200 But the combination of these makes it very, very powerful and very, very complex. 00:07:13.200 --> 00:07:17.860 And so linked open data to my mind is a bit similar in that principle. 00:07:17.860 --> 00:07:22.800 So what we wanted to do in the mining and modeling text project is to bring these three 00:07:22.800 --> 00:07:27.880 things together and so use machine learning to extract information from text, model it 00:07:27.880 --> 00:07:33.940 in linked open data, and use that for literary history, basically. 00:07:33.960 --> 00:07:41.640 So that means we collect information like the one that is at the bottom here, the second 00:07:41.640 --> 00:07:42.900 part. 00:07:42.900 --> 00:07:48.240 For example, bibliographic data, a person is the author of a given work, or content 00:07:48.240 --> 00:07:53.460 related data, like a work is about a given theme, or formal information, like a work 00:07:53.460 --> 00:07:55.860 has a specific narrative form. 00:07:55.860 --> 00:07:57.000 All kinds of statements like this. 00:07:57.000 --> 00:08:00.880 So these are like the patterns or the building blocks, the abstract patterns of the kind 00:08:00.880 --> 00:08:04.520 of knowledge that we're looking for. 00:08:04.520 --> 00:08:12.100 And so basically what this leads to, what we wanted to do is build a Wikidata for literary 00:08:12.100 --> 00:08:13.100 history. 00:08:13.100 --> 00:08:17.260 And so we separated actually in our own Wikibase instance, so we have our own little Wikidata 00:08:17.260 --> 00:08:21.640 just for literary history and actually just for French second half of the 18th century 00:08:21.640 --> 00:08:24.680 novel as a pilot. 00:08:24.680 --> 00:08:29.480 And what this implies, this approach with linked open data, you could call it an atomization 00:08:29.480 --> 00:08:30.820 of literary history. 00:08:30.820 --> 00:08:35.740 So we're breaking things down to the smallest elements, like subjects, predicates, objects, 00:08:35.740 --> 00:08:36.980 and a lot of those. 00:08:36.980 --> 00:08:42.040 And so we're sort of rejecting the grand narrative approach to literary history, where you have 00:08:42.040 --> 00:08:48.380 some idea that literature becomes, I don't know, the novel becomes the genre of the bourgeoisie 00:08:48.380 --> 00:08:50.380 and then you want to show that. 00:08:50.380 --> 00:08:53.020 You have a thesis and then you show how that happens. 00:08:53.020 --> 00:08:55.300 And so we say, no, we don't have a grand narrative. 00:08:55.300 --> 00:09:00.300 We go break everything down to smaller pieces, but then we reassemble this in a model and 00:09:00.300 --> 00:09:04.860 try to find patterns and sort of build literary history from the bottom up again. 00:09:04.860 --> 00:09:08.580 And we'll see how that differs from other approaches. 00:09:08.580 --> 00:09:12.160 And in doing so, we follow some key values. 00:09:12.160 --> 00:09:13.720 This approach is networked. 00:09:13.720 --> 00:09:18.780 So all the information is linked between each other because one author wrote multiple texts, 00:09:18.780 --> 00:09:21.060 one theme is used by multiple texts, et cetera. 00:09:21.060 --> 00:09:25.380 So it creates a network, but also networked outside of our Wikibase. 00:09:25.380 --> 00:09:30.220 So our Wikibase is linked to other Wikibases or to Wikidata, to the National French Library 00:09:30.860 --> 00:09:34.900 and so it is part of a networked system of networks in a way. 00:09:34.900 --> 00:09:36.340 And that's the semantic web basically. 00:09:36.340 --> 00:09:38.620 So we're trying to be part of that. 00:09:38.620 --> 00:09:40.900 We try to be very open about the data. 00:09:40.900 --> 00:09:46.420 So everything is online, everything is freely available, all the analysis are available, 00:09:46.420 --> 00:09:51.460 all the code, all the data, and everything is linked from that Wikibase. 00:09:51.460 --> 00:09:56.100 It's also collaborative necessarily because doing this requires quite a lot of different 00:09:56.100 --> 00:09:59.620 competencies that usually not one person has. 00:09:59.620 --> 00:10:03.620 And it's also collective in that sense that we rely on existing other data sets as well 00:10:03.620 --> 00:10:05.380 that we use. 00:10:05.380 --> 00:10:09.780 And it's multilingual, which I like because I think we live in a multilingual world. 00:10:09.780 --> 00:10:14.860 And because we have this distinction between the identifier and the label, the identifier 00:10:14.860 --> 00:10:20.020 stays the same, the numerical one, but the label can be adjusted to any language and 00:10:20.020 --> 00:10:23.860 you can have multiple labels and the system is set up for this. 00:10:23.860 --> 00:10:26.260 So we like this about it. 00:10:26.260 --> 00:10:32.380 But let's see how this went really specifically in the Mining and Modeling Text project. 00:10:32.380 --> 00:10:36.140 So this is just a very simple overview of the project. 00:10:36.140 --> 00:10:40.500 Basically we had three different kinds of sources that we wanted to use and bring together. 00:10:40.500 --> 00:10:43.140 The first of them is bibliographic metadata. 00:10:43.140 --> 00:10:47.980 So just which author wrote which novel, published in what year, things like that. 00:10:47.980 --> 00:10:53.100 And there is a bibliography of that, of our domain, the French 18th century novel that 00:10:53.180 --> 00:10:55.100 we were able to use. 00:10:55.100 --> 00:10:57.820 And I will show it in a second. 00:10:57.820 --> 00:11:01.460 So this gives us the basis, like all the works and authors. 00:11:01.460 --> 00:11:06.600 Then we also used primary texts, so novels in this case, and I will get to that as well, 00:11:06.600 --> 00:11:07.840 to derive features from them. 00:11:07.840 --> 00:11:14.500 So we did analysis on them and derive information from the novels themselves in a way. 00:11:14.500 --> 00:11:18.120 And the third source of information is works of literary history. 00:11:18.160 --> 00:11:23.600 So chapters in overview works about 18th century French literature, for example, is usually 00:11:23.600 --> 00:11:25.100 a chapter on the novel. 00:11:25.100 --> 00:11:28.800 And so this is another corpus that we used to derive information because obviously they 00:11:28.800 --> 00:11:34.660 will talk about authors and works of that period in such chapters. 00:11:34.660 --> 00:11:40.400 And then bring all that information together in one network of information, basically the 00:11:40.400 --> 00:11:43.000 knowledge network up there. 00:11:43.000 --> 00:11:47.120 So just to illustrate this a little bit, this is from the Bibliographie du Genre Romanesque 00:11:48.120 --> 00:11:51.840 So this is what you see here is one typical entry for one work. 00:11:51.840 --> 00:11:57.280 You see up there the author name Voltaire and then the title Condit Ouloptimisme and 00:11:57.280 --> 00:12:00.080 then all kinds of information and pretty rich information. 00:12:00.080 --> 00:12:02.520 This is not just bibliographic information. 00:12:02.520 --> 00:12:06.920 There are different editions, there are reviews, and something that we were particularly interested 00:12:06.920 --> 00:12:09.120 in because we're not doing book history. 00:12:09.120 --> 00:12:13.840 There's another project using this bibliography for book history, but we're doing literary 00:12:13.840 --> 00:12:14.840 history. 00:12:15.560 --> 00:12:18.880 We're interested in this, what you see here in italics, in the middle of the screen. 00:12:18.880 --> 00:12:22.440 So it gives us information about narrative perspective, Troisiepersonnes, third person 00:12:22.440 --> 00:12:23.440 narrative. 00:12:23.440 --> 00:12:28.760 It gives us information about the narrative locations, Europe and America. 00:12:28.760 --> 00:12:33.520 It gives us the protagonist names, a little bit something about the plot elements like 00:12:33.520 --> 00:12:43.560 voyage, desastres, and also about the themes and even the tone, like satirical tone there 00:12:43.560 --> 00:12:44.560 at the end. 00:12:45.280 --> 00:12:50.360 We digitized the book, we did OCR on it, and then we trained a machine learning classifier 00:12:50.360 --> 00:12:54.240 to pick out these bits and pieces and get them into structured information. 00:12:54.240 --> 00:12:59.520 And that was the basis for the database because that gave us anchors, authors and works, and 00:12:59.520 --> 00:13:05.800 any additional information from the other two sources we would connect to those. 00:13:05.800 --> 00:13:07.480 And it's an almost complete bibliography. 00:13:07.480 --> 00:13:12.960 We will get to the point with the almost because this was done in the 70s and they missed out 00:13:12.960 --> 00:13:16.040 on a few things. 00:13:16.040 --> 00:13:20.320 So the second source of information is the corpus of novels. 00:13:20.320 --> 00:13:25.200 And so we purpose built a corpus for this because not a lot of novels are digitized 00:13:25.200 --> 00:13:26.200 for this period. 00:13:26.200 --> 00:13:30.680 We even had to train an OCR model for this. 00:13:30.680 --> 00:13:35.440 But we ended up with 200 novels encoded in very nice, very simple, but very nice XML 00:13:35.440 --> 00:13:38.420 TEI with some metadata. 00:13:38.420 --> 00:13:40.080 And then we did analysis on this. 00:13:40.200 --> 00:13:43.960 For example, topic modeling to know something about the themes of these novels, named entity 00:13:43.960 --> 00:13:48.520 recognition to know about the protagonists and the narrative locations, like which cities 00:13:48.520 --> 00:13:54.840 are mentioned in the text, and stylometry to know how similar is this novel stylistically 00:13:54.840 --> 00:13:56.280 to all the other novels. 00:13:56.280 --> 00:13:58.160 And so we have data on that as well. 00:13:58.160 --> 00:14:02.680 The corpus was published and described by Julia Röttgemann, and you can follow, you 00:14:02.680 --> 00:14:05.400 can look it up there. 00:14:05.400 --> 00:14:08.560 And then the third source of information is the scholarly literature. 00:14:09.040 --> 00:14:13.240 And here I show an annotation interface for this. 00:14:13.240 --> 00:14:17.960 So this is called Inception to annotate things. 00:14:17.960 --> 00:14:24.360 And so we annotate student helpers annotated these chapters using like following annotation 00:14:24.360 --> 00:14:29.840 guidelines, especially for information on genre and theme of the novels. 00:14:29.840 --> 00:14:33.360 And one of the nice things about Inception is that it's already part of this linked open 00:14:33.360 --> 00:14:38.440 data ecosystem, because you can connect it to any Wikibase instance, our own instance, 00:14:39.320 --> 00:14:43.360 or Wikidata instance, and use it to disambiguate your annotations. 00:14:43.360 --> 00:14:48.200 And you see here, maybe you can see it, there's the bottom one, the yellow one at the bottom, 00:14:48.200 --> 00:14:54.920 it says, Condit, but is it Condit the character in the book, or is it Condit the book? 00:14:54.920 --> 00:14:58.440 So if you look at the context, you will see that it is the book. 00:14:58.440 --> 00:15:03.320 And the annotators can make that explicit by checking the right box here on the right, 00:15:03.320 --> 00:15:06.640 Condit the book, not Condit the character in the book. 00:15:06.840 --> 00:15:11.800 And so we can, because the system knows these are the categories, we can, and these come 00:15:11.800 --> 00:15:16.720 up automatically when you mark Condit, you can simply disambiguate this. 00:15:16.720 --> 00:15:20.120 So this gave us a lot of interesting information already, and we wanted to use this for machine 00:15:20.120 --> 00:15:21.980 learning to scale this up. 00:15:21.980 --> 00:15:23.440 That turned out to be really hard. 00:15:23.440 --> 00:15:26.080 So within the project time, we didn't really get to that step. 00:15:26.080 --> 00:15:31.040 So this is the exception where we didn't just use automatically generated data, but also 00:15:31.040 --> 00:15:33.360 manually generated data. 00:15:33.360 --> 00:15:37.440 To bring this all together, we needed to model it in linked open data. 00:15:37.440 --> 00:15:41.600 And so we started developing a data model for literary history, which is a very complicated 00:15:41.600 --> 00:15:44.600 domain and nobody can really agree on anything. 00:15:44.600 --> 00:15:48.240 But we just said, okay, we need to make a proposal for some fundamental things. 00:15:48.240 --> 00:15:52.640 And actually, nobody can even agree on what is fundamental. 00:15:52.640 --> 00:15:55.840 What is the kind of thing that you want to know about literary history? 00:15:55.840 --> 00:16:01.240 But we just went ahead and tried to find some interesting information. 00:16:01.240 --> 00:16:05.880 So there are modules, it's modular, so you can take out bits and pieces without using 00:16:05.880 --> 00:16:08.840 the whole data model. 00:16:08.840 --> 00:16:14.040 But there are modules on theme, on space, on narrative form, on works, on authors, on 00:16:14.040 --> 00:16:19.760 mapping things, and mapping not in the sense of geographical mapping, but mapping identifiers 00:16:19.760 --> 00:16:23.240 to other resources, and many more, many more things. 00:16:23.240 --> 00:16:29.360 And I will just show you as an example, and don't be scared when you see this, just one 00:16:29.360 --> 00:16:31.160 visualization of one of the modules. 00:16:32.080 --> 00:16:35.800 And there's a lot going on here, but we can just focus on some of the central things. 00:16:35.800 --> 00:16:43.040 So there's a node labeled about, and it has on the top left, it has another node that 00:16:43.040 --> 00:16:44.520 says literary work. 00:16:44.520 --> 00:16:48.000 And on the right, it says thematic concept down there. 00:16:48.000 --> 00:16:54.520 So basically, this is about expressing statements where a work is about a theme. 00:16:54.520 --> 00:16:56.040 Now that sounds super simple. 00:16:56.040 --> 00:16:58.240 Why is it such a complicated image? 00:16:58.320 --> 00:17:04.800 That's because in addition to having these statements, we also have two more elements. 00:17:04.800 --> 00:17:08.520 One is, for example, for theme, we have a thematic vocabulary. 00:17:08.520 --> 00:17:12.560 So a controlled vocabulary, a bit like an ontology, but a bit simpler than an ontology, 00:17:12.560 --> 00:17:15.760 that basically says these are the possible themes. 00:17:15.760 --> 00:17:21.120 And we can have multilingual labels, and we can map topics or annotations to these themes. 00:17:21.120 --> 00:17:25.400 And that's the part on the right where we have this thematic vocabulary, and with the 00:17:25.400 --> 00:17:27.220 labels, etc. 00:17:27.220 --> 00:17:31.700 And the other thing is, so any annotation is mapped to this thematic vocabulary. 00:17:31.700 --> 00:17:36.820 And the other thing is that every statement that says this work is about this theme has 00:17:36.820 --> 00:17:37.860 a source. 00:17:37.860 --> 00:17:40.900 And that's the part here on the top with stated in. 00:17:40.900 --> 00:17:45.140 So anything that has a source, it could come from the bibliography, it could come from 00:17:45.140 --> 00:17:48.340 topic modeling, it could come from the manual annotations. 00:17:48.340 --> 00:17:51.820 So we make this very transparent, and it allows us also to compare information from different 00:17:51.820 --> 00:17:55.340 sources on the same topic. 00:17:55.460 --> 00:18:02.540 Now, the third important feature of this database is that it's linked also to external databases. 00:18:02.540 --> 00:18:05.940 And this is the illustration of how it's linked to Wikidata. 00:18:05.940 --> 00:18:07.260 And it's linked both ways. 00:18:07.260 --> 00:18:08.500 That's the main important thing. 00:18:08.500 --> 00:18:12.460 So authors, works, and thematic concepts, for example, are linked. 00:18:12.460 --> 00:18:15.280 Also spatial concepts like cities are linked. 00:18:15.280 --> 00:18:20.540 And that means that any item like that in our database has Wikidata identifiers, so 00:18:20.540 --> 00:18:24.380 that we know what it is in Wikidata. 00:18:24.420 --> 00:18:27.100 That means we can pull additional information from Wikidata. 00:18:27.100 --> 00:18:29.980 For example, for cities, we don't encode geographical location. 00:18:29.980 --> 00:18:32.540 We just pull that from Wikidata. 00:18:32.540 --> 00:18:35.960 Because otherwise it would be redundant, it would be error prone, you would have to make 00:18:35.960 --> 00:18:37.700 corrections in multiple places. 00:18:37.700 --> 00:18:38.960 So that's one way. 00:18:38.960 --> 00:18:45.880 But we also set up a Memotex ID and registered it with Wikidata and also registered our Wikibase 00:18:45.880 --> 00:18:48.720 with Wikidata as part of the federation system. 00:18:48.720 --> 00:18:54.100 So you can find an author or novel in Wikidata, like the big one, the normal one, and discover 00:18:54.100 --> 00:18:55.940 that, oh, there's a Wikidata ID. 00:18:55.940 --> 00:18:56.940 What do they have? 00:18:56.940 --> 00:18:58.760 And you can pull in our information also from Wikidata. 00:18:58.760 --> 00:19:01.620 So the federation goes both ways, which is pretty cool. 00:19:01.620 --> 00:19:05.200 Also for people to discover our data, because nobody knows that we're doing this. 00:19:05.200 --> 00:19:08.500 But if you are interested in an author and there are very few statements on Wikidata 00:19:08.500 --> 00:19:13.640 and you see, oh, there's another project that has statements on this, this could be interesting. 00:19:13.640 --> 00:19:16.700 The result then is this what we call the Memotex base. 00:19:16.700 --> 00:19:21.820 It's on the one hand just like a Wikipedia page or Wikipedia platform. 00:19:21.820 --> 00:19:24.840 So you can browse and search and click around. 00:19:24.840 --> 00:19:29.340 But more interestingly and more importantly, there's also a so-called Sparkle endpoint. 00:19:29.340 --> 00:19:35.140 So an API to the database where you can write all kinds of queries and ask all kinds of 00:19:35.140 --> 00:19:36.380 questions. 00:19:36.380 --> 00:19:41.940 And this is what I want to illustrate now with an example. 00:19:41.940 --> 00:19:45.340 So we talked about fanfiction and how it's sometimes juicy. 00:19:45.340 --> 00:19:48.300 And that's not a new invention. 00:19:48.300 --> 00:19:56.500 So there is a pretty rich tradition of soft pornographic, erotic, pretty rough pornographic 00:19:56.500 --> 00:19:58.700 literature in the 18th century. 00:19:58.700 --> 00:20:07.140 And there seems to be a sort of structural mapping or concordance between epistolary, 00:20:07.140 --> 00:20:08.640 so writing letters. 00:20:08.640 --> 00:20:11.260 So they are epistolary novels in the 18th century. 00:20:11.260 --> 00:20:14.740 And there's some, you know, they have all kinds of themes, but there seems to be a sort 00:20:14.740 --> 00:20:20.100 of way that libertinage, so which is all of the free thinking, whether philosophically 00:20:20.100 --> 00:20:26.500 or socially, that this goes together well with letters. 00:20:26.500 --> 00:20:34.300 And so this is just an illustration from a late, like an early 20th century edition of 00:20:34.300 --> 00:20:41.020 one of the most famous libertine epistolary novels of the 18th century, The Dangerous 00:20:41.020 --> 00:20:42.020 Liaisons. 00:20:42.060 --> 00:20:47.580 You may have seen the movie with John Markovich and Glenn Close, but the novel is actually 00:20:47.580 --> 00:20:48.620 amazing. 00:20:48.620 --> 00:20:53.140 And so this comes from, this illustration comes from a later edition where you see that 00:20:53.140 --> 00:20:58.260 libertinage and letter writing is very closely connected to each other because you basically 00:20:58.260 --> 00:21:00.380 do it at the same time. 00:21:00.380 --> 00:21:03.060 But anyways, a bit of research on this. 00:21:03.060 --> 00:21:09.900 So is there a libertine epistolary novel after 1782, which is the year of Dangerous Liaisons? 00:21:09.900 --> 00:21:10.900 One author said no. 00:21:10.980 --> 00:21:12.860 There's barely anything. 00:21:12.860 --> 00:21:16.740 The epistolary genre is hardly represented in the libertine novel after Laclou, which 00:21:16.740 --> 00:21:20.700 is a metaphor for this novel. 00:21:20.700 --> 00:21:25.700 And then another scholar said, actually wrote an article on the late libertine epistolary 00:21:25.700 --> 00:21:30.460 novel and speaks about eight different libertine epistolary novels after 1782. 00:21:30.460 --> 00:21:31.460 So what's going on? 00:21:31.460 --> 00:21:33.740 Is there, does it exist or not? 00:21:33.740 --> 00:21:37.420 And on what does it depend? 00:21:37.420 --> 00:21:38.860 And this is what we wanted to investigate. 00:21:38.860 --> 00:21:41.660 So there seems to be this convergence, but when does it happen? 00:21:41.660 --> 00:21:42.820 How does it happen? 00:21:42.820 --> 00:21:48.740 And a database of all this rich metadata on the novel of that time seems to be well placed 00:21:48.740 --> 00:21:50.140 to do this. 00:21:50.140 --> 00:21:58.780 So we try to clarify, or I try to clarify as an example basically, what we can learn 00:21:58.780 --> 00:22:01.300 about this question from our database. 00:22:01.300 --> 00:22:07.180 So first of all, we can query the number of novels that we have in the database for this 00:22:07.260 --> 00:22:12.380 time period, 1782 to 1800, so roughly 20 years. 00:22:12.380 --> 00:22:16.700 And there are 647 novels in the database for that time. 00:22:16.700 --> 00:22:19.780 There are 2000 in total, so that kind of makes sense. 00:22:19.780 --> 00:22:24.620 It's a pretty productive end of the century in a way. 00:22:24.620 --> 00:22:28.020 Then we can also say, okay, but how many of those are epistolary novels? 00:22:28.020 --> 00:22:31.780 Because that's an annotation level that we have. 00:22:31.780 --> 00:22:36.780 And it boils down to 92, 91 novels in query two. 00:22:37.380 --> 00:22:38.380 And what are their topics? 00:22:38.380 --> 00:22:39.380 And what are these about? 00:22:39.380 --> 00:22:40.820 They are candidates in a way. 00:22:40.820 --> 00:22:43.740 They could be about Libertinage or not. 00:22:43.740 --> 00:22:48.140 And so this is one of the queries, the query results that we can get. 00:22:48.140 --> 00:22:52.420 And I can, we can look into the queries together in the discussion if you would like to, but 00:22:52.420 --> 00:22:58.660 this is just a result, a visualization that shows that the epistolary novel is actually 00:22:58.660 --> 00:23:04.060 about sentiments, love, and happiness, about correspondence of course, because epistolary 00:23:04.140 --> 00:23:08.220 novels are always also about writing and reading and sending letters. 00:23:08.220 --> 00:23:09.220 It's like a bit self-reflective. 00:23:09.220 --> 00:23:15.020 Family, virtue, travel, so it's not Libertin at all. 00:23:15.020 --> 00:23:19.460 Except maybe for, at the top left corner you see crime and passion. 00:23:19.460 --> 00:23:20.940 That could be something. 00:23:20.940 --> 00:23:24.900 But actually when we look into it, we see that this is something different. 00:23:24.900 --> 00:23:29.060 Anyway, so this is one of the, one type of information we can get. 00:23:29.060 --> 00:23:31.100 But let's dig a little bit further. 00:23:31.140 --> 00:23:36.380 So how many of those novels, so the epistolary novels of the right period, are actually about 00:23:36.380 --> 00:23:37.380 Libertinage? 00:23:37.380 --> 00:23:39.580 In turns, only one. 00:23:39.580 --> 00:23:44.820 So it breaks down very quickly to just one novel, and it's a novel by the Marquis de 00:23:44.820 --> 00:23:51.060 Sade called Aline Valcourt, which is a sort of compendium of the 18th century novel, one 00:23:51.060 --> 00:23:52.900 of my favorite novels of that time. 00:23:52.900 --> 00:23:55.900 Very long, very complicated, but it has everything in it. 00:23:55.900 --> 00:23:59.020 And it's also epistolary. 00:23:59.020 --> 00:24:02.620 And with the Marquis de Sade, you can imagine that it is also about Libertinage. 00:24:02.620 --> 00:24:03.620 But there's only one. 00:24:03.620 --> 00:24:07.140 So how could Meloncon write about eight of them? 00:24:07.140 --> 00:24:09.500 What happened here? 00:24:09.500 --> 00:24:12.460 So we'll look into that. 00:24:12.460 --> 00:24:17.180 There are some non-epistolary Libertine novels, 14 even, of them at that time. 00:24:17.180 --> 00:24:19.820 So apparently it's not like Libertinage is dead. 00:24:19.820 --> 00:24:25.860 It's just that this combination seems to be gone, or might be gone. 00:24:26.020 --> 00:24:30.460 And for example, just to show what we can visualize, these Libertine novels of that 00:24:30.460 --> 00:24:34.340 time period, we can show where they take place, where the action takes place. 00:24:34.340 --> 00:24:40.060 And this is the second, like, query six, I think, just the output. 00:24:40.060 --> 00:24:41.060 And it's a bit hard to see. 00:24:41.060 --> 00:24:42.060 And again, we can look into it. 00:24:42.060 --> 00:24:44.620 This is interactive in the database. 00:24:44.620 --> 00:24:46.780 Here it's just static. 00:24:46.780 --> 00:24:51.900 But we can see that it's set in different places of Europe, and also in one of the novels 00:24:51.900 --> 00:24:55.540 is set in Africa as well. 00:24:56.220 --> 00:25:00.860 In this case, this is like the result of a federated query, where we pull out the information 00:25:00.860 --> 00:25:06.500 and narrative location that we built, and combine it with the latitude and longitude 00:25:06.500 --> 00:25:12.300 information from Wikidata, and then visualize it in the system. 00:25:12.300 --> 00:25:14.340 But let's get back to that question. 00:25:14.340 --> 00:25:15.460 So what went on? 00:25:15.460 --> 00:25:19.580 Why is it so different for Meloncon and for our database? 00:25:19.620 --> 00:25:26.260 And it turns out, so I looked at the candidates that Meloncon wrote about, what are these 00:25:26.260 --> 00:25:27.260 novels? 00:25:27.260 --> 00:25:30.460 And so some of them are not marked as epistolary in our database. 00:25:30.460 --> 00:25:32.340 So it's a formal feature, you would think. 00:25:32.340 --> 00:25:33.340 It's simple. 00:25:33.340 --> 00:25:34.340 It's yes or no. 00:25:34.340 --> 00:25:36.020 But it's not, because they are mixed forms. 00:25:36.020 --> 00:25:38.640 So there's first-person narrative with some letters. 00:25:38.640 --> 00:25:41.340 And we didn't mark that as an epistolary novel. 00:25:41.340 --> 00:25:43.600 We just marked it as mixed. 00:25:43.600 --> 00:25:44.740 So it's not in there. 00:25:44.740 --> 00:25:46.340 But he said, yeah, there's epistolary. 00:25:46.500 --> 00:25:47.000 That's fine. 00:25:47.000 --> 00:25:50.460 So I will keep that in my corpus, basically. 00:25:50.460 --> 00:25:54.420 Some of them are not marked as libertine in our database. 00:25:54.420 --> 00:25:57.500 But he uses them as examples of libertine novels. 00:25:57.500 --> 00:26:03.780 And I think that's simply because libertinage is not an easy concept to cut into pieces. 00:26:03.780 --> 00:26:06.980 And we actually have an item, libertinage. 00:26:06.980 --> 00:26:10.460 But we also have related topics that are actually included. 00:26:10.460 --> 00:26:15.580 And so you can actually play with the breadth of the concept in a way, by saying, I will 00:26:15.580 --> 00:26:22.540 only look at novels that have libertinage as keyword, or also sexuality or eroticism 00:26:22.540 --> 00:26:24.700 or other things like that. 00:26:24.700 --> 00:26:27.720 So you can vary the precision. 00:26:27.720 --> 00:26:32.500 But even with a rather wide definition, we don't have as many novels marked with this 00:26:32.500 --> 00:26:33.500 topic. 00:26:33.500 --> 00:26:36.820 So it's a question of definition and of the sources. 00:26:36.820 --> 00:26:40.380 And then something that was a bit shocking to me, two novels that Melonçon talks about 00:26:40.380 --> 00:26:42.260 are missing from our database. 00:26:42.260 --> 00:26:44.540 And that can happen. 00:26:44.540 --> 00:26:48.660 And it turns out these are two novels that were re-edited after the bibliography was 00:26:48.660 --> 00:26:49.820 published. 00:26:49.820 --> 00:26:52.400 So we probably should have edited in the meantime. 00:26:52.400 --> 00:26:56.100 But it makes sense that they didn't know about them yet. 00:26:56.100 --> 00:26:57.780 Although they have a lot of unedited novels as well. 00:26:57.780 --> 00:27:00.740 But somehow these must have been rediscovered later. 00:27:00.740 --> 00:27:04.420 Anyways, I will come to the conclusion now. 00:27:04.420 --> 00:27:08.100 So basically what we see is that we have multiple perspectives in the database. 00:27:08.100 --> 00:27:12.220 But Melonçon has yet another perspective. 00:27:12.220 --> 00:27:13.220 So I conclude. 00:27:13.220 --> 00:27:14.220 Yes, wonderful. 00:27:14.900 --> 00:27:17.940 I think I will just about make it. 00:27:17.940 --> 00:27:21.940 So some challenges and some opportunities of this approach. 00:27:21.940 --> 00:27:27.040 One of the challenges is that there's necessarily a certain complexity reduction when you try 00:27:27.040 --> 00:27:32.620 to take literary history and all of its diversity and you want to express it in triples. 00:27:32.620 --> 00:27:33.620 That's normal. 00:27:33.620 --> 00:27:34.620 But that's inherent in modeling. 00:27:34.620 --> 00:27:38.980 So there's a trade-off between this reduction in complexity and then what you can do with 00:27:38.980 --> 00:27:39.980 the data. 00:27:40.740 --> 00:27:45.700 The upsides of having the data in this shape really outweigh the limitations of the modeling. 00:27:45.700 --> 00:27:51.780 But we have seen that libertinage and even narrative form are not clear-cut categories. 00:27:51.780 --> 00:27:56.460 There's a lack of consensus, as I said, on relevant statements and how to model the domain. 00:27:56.460 --> 00:27:59.180 And that's just a challenge that's not going to go away. 00:27:59.180 --> 00:28:02.940 It's not something you can solve with technical solutions. 00:28:02.940 --> 00:28:05.740 Then there's this federation thing. 00:28:05.740 --> 00:28:09.660 The vision of linked open data and of semantic web is federation. 00:28:10.340 --> 00:28:12.700 It's not to have one central database, everything in Wikidata. 00:28:12.700 --> 00:28:14.980 It's these distributed resources. 00:28:14.980 --> 00:28:17.820 But it's actually really hard to get it to work. 00:28:17.820 --> 00:28:21.980 It took us a long time and towards the end of the project was when we actually managed 00:28:21.980 --> 00:28:23.140 to do this. 00:28:23.140 --> 00:28:27.740 And then sustainability is a big challenge because who's going to take care of our data 00:28:27.740 --> 00:28:29.540 when the project has ended? 00:28:29.540 --> 00:28:33.660 So we have another project so we can sustain it, but it's not really a good solution. 00:28:33.660 --> 00:28:37.980 We could post everything on Wikidata and parts of the data we have pushed to Wikidata because 00:28:37.980 --> 00:28:40.260 we think that's too big to fail in a way. 00:28:40.260 --> 00:28:44.940 And of course we have exported the raw RDF data to Zenodo, but there you can't really 00:28:44.940 --> 00:28:50.660 do anything with it, at least unless you analyze the RDF directly, which you can do, of course, 00:28:50.660 --> 00:28:52.900 but not everybody wants to do that. 00:28:52.900 --> 00:28:55.940 But there are, of course, also many opportunities. 00:28:55.940 --> 00:29:01.940 It's really a great way of linking these heterogeneous data from different types of sources and use 00:29:02.020 --> 00:29:08.220 semantic modeling to bridge differences in granularity, for example, in terminology, 00:29:08.220 --> 00:29:09.980 in language. 00:29:09.980 --> 00:29:14.820 It's a way to model differences of perspective on data because we have these sources, so 00:29:14.820 --> 00:29:19.500 we can have contradictory statements or statements of different granularity on the same thing. 00:29:19.500 --> 00:29:25.100 But that means we are modeling not facts, but perspectives or just statements, things 00:29:25.100 --> 00:29:27.620 we found. 00:29:27.660 --> 00:29:32.540 Like somebody said in the conference at the beginning of the week on linked open data, 00:29:32.540 --> 00:29:36.300 that history is never wrong. 00:29:36.300 --> 00:29:38.340 It's just what it is. 00:29:38.340 --> 00:29:41.540 There's transparency and knowledge production in this way because everything is sourced 00:29:41.540 --> 00:29:45.900 and you can click through to the GitHub repository with the topic model and the corpus and the 00:29:45.900 --> 00:29:47.380 code and everything. 00:29:47.380 --> 00:29:51.980 It's multilingual, which I think is really important, and we can avoid redundancy, as 00:29:51.980 --> 00:29:55.700 I said, by reusing external resources as well. 00:29:55.780 --> 00:29:59.500 Because this is so wonderful, we want to continue doing this, and so we started a new project 00:29:59.500 --> 00:30:05.620 called linked open data in the humanities, Loading, where we work on scholarly digital 00:30:05.620 --> 00:30:12.140 editions of correspondences, for example, which you can model as networks, on bibliographical 00:30:12.140 --> 00:30:15.340 data, on the digital humanities themselves. 00:30:15.340 --> 00:30:21.660 This is a project on the institutional history and institutional landscape of digital humanities, 00:30:21.740 --> 00:30:29.140 so you see that it's slightly biased towards Europe and the east coast of North America. 00:30:29.140 --> 00:30:35.180 Another project on bibliographical data and my favorite project on wine labels, such as 00:30:35.180 --> 00:30:40.940 these that we are digitizing and networking and annotating and mapping, of course, and 00:30:40.940 --> 00:30:43.140 showing. 00:30:43.140 --> 00:30:44.140 And that's it. 00:30:44.140 --> 00:30:44.900 Thanks for your attention.