Theoretical Perspectives I (March 14):
Paul Caton, “Towards an Ontological Model of Text” (video)
[Fotis Jannidis] As probably most of you have already heard Claus Huitfeldt can’t come, so for this session we will have three speakers. Go ahead!
[Paul Caton] I was really pleased the other day to get Julia [Flanders]’s email where she said, perhaps anticipating some over-long talks, “Well, we’re not looking for highly polished presentations of finished research; it’s more about provocative ideas and things. Try and keep it short and sweet.” And I thought, “Oh, perfect, this is great.” I’ll explain why this is perfect. And unfortunately now that Claus isn’t here that falls to the ground.
Originally when I agreed to do this it was on the assumption that I’d be coming out of or at least would be well into a period of research and that in fact I would have some results, some conclusion to give to you. And that in fact I would be addressing you from a high point. I would have climbed a mountain and I would be on a high point and I would be telling you what I could see, and describing it all in beautiful detail, and of course you would all be very impressed. And that is not going to happen. I am not at a high point. I am at a low point, not morally, but research-wise. And let me show you what I mean. Each of those five pieces [on slide], each of those five strings describes something I have been working on. And each of those could have been a high point that I ascended to and could have been telling you all about what I could see in great clarity. In fact I could have been on a couple of them at once, or I could have been on an intervening ridge like the peak at Darien, able to see two oceans at once. Instead of which, I am not at the high point of any of those. In fact if I had to locate myself, if you’re interested in organizing a search, it probably would be somewhere in there, which is not a peak, it’s a valley. A heavily forested valley.
Let me just quickly, very briefly explain to you what these five things are. “Model of transcription”: this is an area I’ve been interested in since Michael [Sperberg-McQueen] and Claus’s [Huitfeldt] paper on trying to organize a formal model of transcription, and I sort of followed along after their paper rather like a child going for a walk with its parents. They were kind of striding along ahead interested in getting on and I was sort of lagging behind and sort of noticing things and saying, “Dad, Dad, Mom, look at this! This is really interesting.” And they were tolerant as good parents are but basically saying, “Come along, we’ve got places to go.” But I’m still interested in that. I can’t leave that alone but I’ve had to put it to one side. So I’m not anywhere near a height on that.
“Meaning of text” is something that I’ve been working on for a couple of years. You may have seen a couple of presentations that I’ve done relevant to that. “Model of handwriting” comes out of some work I’m doing as part of my job at DDH [Department of Digital Humanities (King’s College London)], working on the DigiPal or Digital Palaeography Project of which Peter Stokes is the principal investigator. “Model of the domain of medieval handwritten documents” is also related to DigiPal and the world of Ben Jonson as mediated through the Cambridge Edition refers to the fact that, as some of you probably know, Cambridge University Press is about to publish a new complete edition of the works of Ben Johnson and DDH will be responsible for the digital edition. So I’ll show you some slides that relate to each of these five things and explain how I got lost. And let me tell you, I didn’t get this lost overnight. I took five years of hard work to get wherever the hell I am today. This wasn’t an easy thing. So some of you will have seen this before , this is a diagram, a model if you like, that came out of my work that picked up on what Michael [Sperberg-McQueen] and Claus [Huitfeldt] had been doing and it was just adding some stuff that I thought was necessary to it. This is something that I’ve had to put aside. I’m not finished with it yet and some of the work on the other things that I’ve been doing are going to feed back, I hope interestingly, into this. But I just haven’t gotten to do that yet. But that’s where I was with transcription.
This is something I should have done last year actually at Stanford and I don’t know why I didn’t do this because it would have saved me a lot of explaining. But two years ago at DH [conference] at London, at King’s I was just expressing some disquiet about the fact that we use certain terms in Digital Humanities and they don’t seem to be ever really fully defined or explicated anywhere and it was just bothering me and I thought, “We should really pay attention to the terms we’re using and try and figure out exactly what it is we mean when we use them.” And then I thought, “Well if you’re going to stand up and say that then it kind of is beholden on you to do some work along those lines yourself.” So last year at DH I gave this talk and this is a diagram of the approach where I decided, I was concentrating on the term “text” because that seems to me one of the most interestingly problematic of the terms that we use. The approach that I decided to use was, I’ll take a core instance, something that obviously, something—let me rephrase this carefully: Something to which the term “text” obviously applies in humanities because they do so overtly and with great authority, which is the scholarly print edition. If I take that as a core instance and I say, “Well, this seems to have certain characteristics, that every instance of this core instance has certain characteristics,” there must be a point at which those characteristics no longer apply. So that would then be the boundary of what it is to be something to which the term “text” can apply. So if I sort of measure that distance or I come up with some marginal cases and sort of measure if those characteristics still pertain or not then we can come up with the boundary and say beyond that this is not text or “a text,” which is another distinction I was interested in following.
So that was as far as I got with that and I gave the talk at DH and then I sort of reshaped and slightly reworked the paper for submission to the journal for the special issue. And I got some good comments back from the reviewers, very useful. I’m working and reworking the paper based on the the good feedback I got. But one of the comments I got disturbed me not because I thought it was wrong, but well let me—this is a sentence from the paper I gave at DH. This was my conclusion where, I’ll let you read it as I’m talking. [Text of slide: Being “a text” is a status we give some text in a particular context and at our choosing. In this sense “a text” is, as [Allen] Renear and [David] Dubin say in a somewhat similar context, “a matter of contingent social/linguistic circumstances” (2007 p.8) and is thus – as they similarly concluded about three of the four FRBR Group 1 entity types – not a type but a role. In other words I suggest that being “a text” is not what [Nicola] Guarino and [Chris] Welty would term a rigid property of any instance of text in its mass noun sense. A good deal of ontological work needs to be done, however, before this can be asserted with confidence.] Essentially what this is saying is text is a floating signifier. Right? It’s just there in the air in digital humanities and when we feel we need it we grab it and pull it down and slap it on to something. Notice that last sentence there: “A good deal of ontological work needs to be done before this can be asserted with confidence.” Now I actually left that sentence out of the version that I submitted to the journal but the comment that disturbed me—and it was a perfectly reasonable one—was that the reviewer said, “Well I really like the way the argument builds up and there’s lots of interesting examples but then it comes to this conclusion and it’s kind of a let down. It’s kind of lame.” And what disturbed me about that was that I actually thought that was the most exciting conclusion that came out of this, the fact that it’s really hard to pin down a meaning for a text, the fact that it is this floating signifier. I mean to me that seemed like a really knotty thing to deal with but I could see that from another point of view that was actually perhaps of no interest to people at all. That it was just so obvious that how could this possibly be an interesting conclusion, which I thought was fair enough.
But what I found particularly interesting about that was a kind of tension between, for want of a better phrase, I’m going to call it a taxonomic imperative where we tend to want to drive down through a taxonomy to reach a terminal leaf. Whereby we can say “I have distinguished this thing from other things that may be related to it.” But I have got to this point where I can finally say, “I get down here and it goes no further. This is it.” So there’s that. And you can see that driving down from some grandparent node all the way down to a final terminal leaf: “Here I am, I have arrived.” And we sort of like to slap labels on our terminal leaves. But the problem with something like “text” is that it spans across multiple terminal leaves and the tension there between that taxonomic imperative to drive down and find something, distinguish it and label it, and the tendency of natural language to spread itself, to smear itself across is something that I thought was really interesting and it made me realize that more work was needed.
Coming out of that work then on the meaning of the word “text” and the way we use it—two paths looked kind of interesting to me. One was the ontological path, which is instead of saying we have this core—the scholarly print edition and these rather vaguely delineated characteristics, I thought what I should do really is formally specify what is a core instance, what are the characteristics. So that pushes me towards ontology. And then the other thing is to come up with some way of dealing with the relation between the drive to systematize and that taxonomic imperative and the problem of the fuzziness of natural language. So that was another thing, but sometimes you have to put these aside because you have a day job. And part of the day job involves the digital edition of the Cambridge new edition of the works on Ben Jonson. Now the Cambridge edition is vast. It’s been in production now for well over a decade and the digital edition is not exactly an afterthought, not an add-on, but the digital edition is basically the print edition made digital, plus other stuff. Stuff that they simply couldn’t fit into the print edition because it would have been too monumental. So there’s a limited amount of time and money that can be spend on the digital edition so basically what we’re doing is, to put it bluntly, data modelling on the cheap. We do need some entities. We’re going to try and discover some entities in the text through computational means. We’re going to try and find people and places just through automated processing. But when you’re talking about an author like Jonson—when you’re talking about a comprehensive edition you’re going to have lots of strings that refer to different kinds of thing with the same string. For example, the string “Volpone” can refer to many different kinds of thing. It can refer to—and some of these terms you’ll recognize, or they’ll be very similar to terms you know from, say, FRBR—say it can be a work or it can be a version of a work or it can be a realization of a version. I wanted to come up with ways to distinguish the kind of things that might be being referred to by particular strings. So we decided that we would use EATS, which is the Entity Authority Transaction Service. EATS is not a data modelling tool. EATS is really all about a place to store information, to store disambiguating information about entities. But you can twist and pervert it into a poor man’s data modelling tool if you’d like because you can declare your entities, you can declare what types of entities you have and you can declare relations between those entities. So I thought, well, as an early attempt at doing this I’m just going to look at the data and come up with the entities that I want and then specify some relations, and then maybe later on I’ll either translate them to an existing schema or keep them as they are, and just say “This is the same as something from FRBR or FRBRoo or CIDOC-CRM or whatever.”
So here are some early diagrams sort of mapping out entities and relationships, and you can see there we’ve got Ben Jonson as a person and we’ve got Every Man Out of His Humour as a work and then we’ve got a quarto edition version and so on. All very standard kind of entity relationship mapping. And there’s another one, similar kind of thing—we’ve got an event entity. And here’s one of the early EATS records that I created just as part of testing this process and we’ve got an entity type “event”— “Jonson marries Anne Lewis.” Okay, fair enough. And then we’ve got another event entity—”Jonson kills Gabriel Spenser in a duel.” And then I thought I wanted to start mapping this on the diagram too. And here I started having to think and that’s always a bad sign. The more you start thinking about this stuff the more difficult it gets. Here for example, this goes back to something that Andrew [Ashton] was saying in the panel discussion, I was really interested in the relation, not just the entity, the person entity, I mean although they’re fairly standard, but I’m interested in what kind of relations do we need in this. What is the relation between a person and an event? So here for example I’ve come up with “a person has experience of an event.” But I’ve also got a personal relationship between Ben Jonson and this actor Gabriel Spenser. Ben Jonson and Gabriel Spenser had a duel and Gabriel Spenser died. In the same way that you have, say, person marries person, I’ve got Ben Jonson kills Gabriel Spenser. It just seemed though the more I thought about it that seemed like a weird relation between two people. Then I thought maybe this is better . We’ll lose that personal relationship, we’ll take that out and we’ll just have the relationship, the fact that Ben Jonson killed Gabriel Spenser now just becomes mediated through the event that they both have experience in the event. Which is plausible but the problem is the more I thought about this, the more I thought about, “Well, there seems to be two events there in a sense,” or “Where do you draw the line at an event?” Is it that they have a duel as one event and Gabriel Spenser gets wounded as another event and then he dies as another event? What is an event?
And as soon as you start asking the question, “What is X?” you’re doomed because now the gravitational pull of ontology starts sucking you in because you think, “Well there must be smart people who figured out what an event it.” So then you start going to the ontology books and yes, everybody discusses an event but they’re all subtly different so then you go further back trying to find out. An event sounds like a natural category, right? What are natural categories? And then it turns out, well yes they all have similar natural categories but they’re not all the same and they don’t organize them, they don’t organize them the same either. I got stuck in this question about “Does it make sense?” What is the relationship between two people? Should a relation be something that is just a long-term relation? Is an immediate happening like somebody sticks a sword in somebody else and kills them, is that a relationship between two people? But if it’s not then what is the difference between marrying somebody or killing them?
I won’t go any further down that line. Anyway, the moral of the story is these are all reasons for why I’m lost. These are all reasons why I start down or start up a height intending it to be a pleasant hike that will take me to somewhere where I can give you the benefit of my wisdom and I end up lost. This is why. [20:48] It’s hard to do just a ‘little’ data modelling. Even though you start off with the best intentions it’s really hard to stop at any point. Obviously “don’t reinvent the wheel” is a good principle, right? I mean people have spent hundreds of years writing ontologies and modelling entities and entity classes, so don’t reinvent the wheel but don’t assume either that somewhere back down the line there’s a standard wheel that everyone agrees on. Simple entities, it turns out, are never that simple.
Complicating things even further is DigiPal. DigiPal is a fascinating project. The idea is to come up with an extensive database of instances of medieval scholarly handwriting. Particular databases of letter forms so that, the idea that Peter [Shields] has is that he wants to give some kind of objectivity to the way that features of handwriting are described because medieval scholars all know what they’re talking about but they all talk in slightly different terms and they all use slightly different descriptors. And the idea is to try and get some kind of uniformity in description of letter forms. So as part of that Peter wanted to concentrate on modelling handwriting: what components of handwriting and what goes on in handwriting. And one of the things that I was tasked with doing was thinking about how that fits into a wider world of documents, writing, texts, books, et cetera. So being a bit of an idiot I decided to create an ontological framework that would slot Peter’s model when he finalized it, would be able to slot into a model of everything to do with the things that we study. And you can see sort of the general outline of the thing being formed here and it just got more and more complicated. As I tried to define what things were, I just kept having to add more and more classes and categories and I thought this can’t be right. So eventually time ran out and I had to give up on that because it was just unfolding like a kind of Mandelbrot curve. It was ridiculous. So this is the current model. We are actually going with sticking with the relational database. This is the UML model. It’s pretty much final. I can’t say I’m completely happy with it but in the end work had to get done so this is the one we’ve got. Now the really interesting part about this is the part that Peter’s been working on, which is if the vertical column of blue squares starting with ontograph up at the top and going down to graph. And what that is basically is the descent from the most abstract conceptual division at a level with no form. In other words the division between, say, an A and a B without giving either of them any form right down to having something actually written on a piece of parchment. That’s what that is.
What’s really interesting about that model is that I think it’s going to have a great deal to do with the model of transcription, but I just haven’t figured how to plug it in yet but I think it’s going to be really important. And the next step is to try and somehow work all these things together and figure out somewhere how to bring in the ghost that departed which is text because everywhere you go it’s there but it’s never graspable. It’s never really there. I wish I had been counting the number of times people have used “text” or “texts” in the discussion this morning. It’s just part of our vocabulary. But what kind of thing is it? For example when I was trying to work out the basic entities for the Jonson edition, obviously you’ve got fine distinctions there between the various stages of the publication of writing process. And the original FRBR class one entity type simply is just inadequate to deal with that. So that drove me to FRBRoo—which is the object oriented version of FRBR— Allen [Renear] is on the advisory council drafting committee or something. FRBRoo is a terrific piece of work because it actually addresses a great deal of the deficiencies of FRBR class one entity types. But the text isn’t there in FRBRoo; it’s just not as an entity. But if you read the spec which describes the classes that are in FRBRoo, “text” is used many times as if we just all know what it means. And in a sense of course we do all just know what it means because when people used it this morning no one stuck their hand up and said, “What the hell do you mean by that? What are you talking about?” But I’m really interested in this phantasmal thing that permeates our discourse but doesn’t seem to be something we can really easily pin down. And I’d be delighted to be the first person this morning to comment on the wonderful carpet we have, which as I would say contains text. I don’t think it contains a text although you could make a text out of this carpet. And I would love to see Michael [Sperberg-McQueen] use his model to try—how would we transcribe this carpet? I think it could be done but, is it going across the walls? Is it going up and down? It’s an interesting question. So this is where I am. You can see I’m completely lost. I’m floundering around but I think there are ways to go. I’ve decided that there are things that will help me although they’re not easy although they’re not easy because it’s a question of stepping out of our own comfort zone into other people’s disciplines, which of course is A, resented by them and B, is an awful lot of work on your part. So I’m open to collaborative work here.
So here’s where I think I can get out. I’m splitting myself into three people. One person is going out through scientific ontology. I’ve delved briefly into fundamental ontology and it’s a scary place that deals with propositional logic and not much else. It’s like some weird Doctor Who episode where there’s just a white screen and a person and nothing else and they ask questions like, “Why is there something rather than nothing?” And I just want to know what an event is so that’s too scary for me. I’m not going there but the scientific ontology I think will help. The other way out is through set theory, I think. And Wendell [Piez] will know [brief loss of video] “Wendell, I’m sure the answer lies in set theory.” And he emailed me back the equivalent of a pat on the head and said, “You’re probably right. Go ahead.” But I think using set theory in conjunction with those levels you saw in the diagram , the levels between ontograph and graph is going to be interesting and may also help drag my account of transcription back into line with Michael’s and Claus’s where they’ve moved on to trying to deal with the level of an organization rather than a simple stream of tokens. And the other way out I think is through Wittgenstein and people like him, which is the philosophy of language, because I don’t see how else to try and deal with this business of the tension between the floating signifiers that just want to spread itself over everything. And of course we’re very happy to let it do that because it’s convenient for us. And that drive to distinguish something and say what it is, to pin it down and say, “I’ve isolated you now.”
And that’s it.
[Fotis Jannidis] We can take a moment to answer some questions now and then after the second talk we can have a general discussion.
[Jan Christoph Meister] Two observations, I don’t know if they’re really questions-—or one is a question at least. What about ideal types? Why don’t you try and describe not in terms of the five or six characteristics of something and make that distinction between what’s at the border, what’s beyond the outer perimeter of your definition? We know that an ideal type definition of any phenomenon is always a mix of N parameters and some may be required and some may be contingent. So that’s one question. And the second, as for events—I don’t know if I should say it. I will say it. I know somebody who wrote a book on this. It happens to be me. And I came to the conclusion that it was completely nonsensical to think of events as something that exists. An event is something that you describe as an event or define as an event. So it would matter if it was really as to what the definition is going to be. And in certain domains there are valid descriptions and definitions of what an event is and whether it has an existence. So an event, to look for events in the world doesn’t make sense to me.
[Paul Caton] I’m happy to hear that.
[Jan Christoph Meister] You were right to have trouble.
[Paul Caton] I’m sorry?
[Jan Christoph Meister] You were right to have trouble.
[Paul Caton] Well I think I was right to have trouble but that doesn’t get rid of the need to deal somehow with changes in states of affairs. Right?
[Jan Christoph Meister] Well that’s one description under which an event can occur. It happens to be the one that I favor myself. And that’s true Wittgensteinian approach that you’ve taken. But I think it’s a waste of energy to look into events as real live moments.
[Stefan Gradmann] Yeah, I’m so glad you did this talk because it’s just really—
[Paul Caton] Sorry Steve, I can’t [hear]—
[Stefan Gradmann] I’m so glad you gave this talk, from beginning to end. I think questions like “What is an event?” and “What is a text?” and I would go further, “What is a tree? What is a game? What is a door?”—I think these questions are philosophical nonsense. And I notice that your trouble is not your inability to write systems. Because in terms of writing a system, when you try to write a system in which people can locate duels, you can do that successfully. Your trouble arises because you think that there should be some kind of relationship between that kind of functional purpose and the fabric of the universe. You think that there should be some correspondence between that act of making something that sort of works in the world and ontological reality. And when you go through these three clouds and get to Wittgenstein I can hear the sigh in your voice as if you’re jumping off the cliff if we just jettison this entire ontological project and say, “oop.” Text is the sum of the instances in which that word is used in the language. End of story. Now we’ll get back to writing the system that can do something for people in the world. I feel, I’m listening to this and wondering is your problem a real problem? Because to me it sounds like it’s a psychosis. I say that not to insult you but you’re getting trapped in a philosophical conundrum and all I can think of is why not decide to stop having that conundrum? Because solving that conundrum, well you would have to convince me that solving the conundrum would make for better functional, better programs in the world, absolutely no difference about it.
[Paul Caton] Oh, I agree. And there has to be a point in the practical real world where you just say, “okay, that’s it.” I’m going to go with—
[Stefan Gradmann] What other point is there?
[Paul Caton] Well, call it a psychosis or call it what Wendell [Piez] said about himself earlier, an obsession. I can’t leave it alone. I can’t leave it alone. (35:53) I’m actually fascinated by the fact that we use text a lot and we can’t define it. Or we can define it in twenty different ways and they all serve the purpose. I’m sorry. That fascinates me. There comes a point where I have to divorce my pursuit of that from needing to get projects finished. But I also can’t resist the temptation to be the man who finally joins them and achieves lots of fame.
[Maximilian Schich] I think there’s something from out of this contention or tension we have because we’re all interested in these mobile parts . And then the problem is one could be interested in two of them. And there’s so much other things in the texts one could look at things like an event. And then once we start and describe a whole variety of these things then we run into this problem by the end at, at 3:00 at night we dream about describing the universe. But I think that on the other hand there’s another functional thing. For example, Freebase has a very interesting different kind of way of describing what is often called an event, called compound value-types which is basically an empty node and it brings together say a person, a time range, a war, a killing of another person, or whatever, a person, a location, something like that. But the node is empty and that’s the interesting thing. I think the invention of the compound value types type in ontology is a little bit like the invention of zero in mathematics because we can do much more with these zero nodes which don’t have a name or neighbor or anything except—
[Paul Caton] So it’s not a thing in itself. It’s a thing that comes into existence when other things are in particular conjunction.
[Maximilian Schich] And so after, if you call it an event or not, is something you can do.
[Paul Caton] I should say the jewel was just, it was just one of those little, it was just the snag that stuck to my smooth gliding into an easy definition of things, finish the task and get on with my life. It was just that one little thing that hooked the sweater and then the entire sweater began to unravel. And I wrote, Wendell remarked earlier, you know what’s interesting is when the model fails or when the model that you thought was going to be simple starts getting more and more complicated in a recursively nasty way.
[Elke Teich] I’d just like to suggest another area you may want to look at.
[Paul Caton] Oh God. I’ve only got one life to live.
[Elke Teich] It’s completely pragmatic. It’s got nothing to do with ontology, it’s not complicated, and that’s data mining. So instead of trying to construct an ontology you just have a set of features that you looked at in the past. It’s just a list of features. And you characterize the text in your collection by these features or by these attributes or particular values and then on the basis of these aggregate value pairs you just capitulate how typical the text is given, okay your ideal text has all these properties, that’s how the text is defined no matter how these categories are ontologically. And then you test a text for how typical it is given that this is your typical text. And you can apply all kinds of methods from data mining like automatic classification or clustering and just see what happens. And it can be very instructive just to do that. So that’s how your text—
[Paul Caton] What kind of data are you talking about mining here?
[Elke Teich] Well that would be the features, the attributes, the properties you assign to the text.
[Paul Caton] Right.
[Syd Bauman] The problem is what kind of data do you look at to find those kinds of things?
[Elke Teich] Well, your text.
[Paul Caton] Right. If it was a simple linguistic thing that would be easy but I’m quite not sure how to, I mean I see what you’re saying, I’m just not sure how to apply it because the sorts of characteristics I’m looking at are not amenable to that kind of analysis.
[Fotis Jannidis] The way you build your categories—would it be easier to take a step back and say okay, the tree, the hierarchy you showed us is just one way to build plots in the traditional way, but you can say, “I’m using a prototype approach,” instead of building classes by specific features. And then say—and then you would have a totally different outcome actually describing what a text is. And many of the problems you were talking about like the smearing of the features, they are actually artifacts by the approach you would take, how to build your approach. So I think it’s really the way, the meta-tools you use here. They create the problems you then are left with here. At least in the event case and in the text case probably too, as far as I know with the text discussion.
[Stefan Gradmann] I’d like to thank you as well for this talk, I’d like to take up your remark once again about the empty value nodes. In RDF they have a similar concern which is that like you’ve noted. That is something that has no real or actual correspondent entity in so-called reality. Actually the interesting part about this is that it comes into existence by modelling. So the like node is there in the data […] later. You cannot model data that way. Create an entity and then try to see where it relates to the world. So it’s not about data modelling. It’s about conceptual modelling and later eventually binding that to real data and entity modeling. So I think it’s introduced a new […].
[Maximilian Schich] Just another comment, I think data mining is one way how you can extract clusters but the clusters are often how you model them. So you have exactly the same problem as if you start to think about [unclear, someone talking over] So it’s not like data mining would be better than the modelling the first time if you have a lot of [unclear, someone talking over] overlap…
[Fotis Jannidis] [unclear, someone talking over] you have a specific approach to building a class, saying, “I need classes […] and overlap, which is probably not the way we construct our world. And we are talking basically about [construction?]. Not about […] there are no events […] texts are not out there […]
Pingback: Knowledge Organization and Data Modeling in the Humanities: An ongoing conversation | datasymposium