Balisage Paper: There are No Documents

Introduction

Last year at Balisage (2009) we explored the claim that documents, in the sense of repeatable symbolic expressions, cannot undergo any genuine change or modification[renear09]. This is a claim obviously in direct contradiction to the common assumption that documents routinely undergo a wide variety of changes, such as revision, editing, and reformatting. The modification of documents is in fact an integral part of publishing workflows and so at the heart of many activities of professional interest to Balisage participants. Not surprisingly one comment on last year's presentation was: "You are telling me that my profession is impossible".

Our analysis of the claim that documents cannot be modified took the form of identifying and evaluating possible responses to this inconsistent triad:

Late this spring (2010) we were surprised to realize that our survey of possible responses to the document modifiability puzzle had overlooked one response: There are no documents. In the present paper we explore what might be said for this response as a solution to the modifiability puzzle. We also explain how this response escaped what we believed at the time to be a comprehensive survey of solutions.

[This is the pre-conference version of a Balisage "late breaking" paper. These are early provisional reflections, begun recently, and put in written form in order to elicit discussion at the 2010 conference. We are looking forward to corrections, criticisms, and suggestions, which will inform the final version.]

The Arguments

The word "document" has, of course, many senses. Our informal characterization of the sense intended in "Documents cannot be modified" is repeatable symbolic expression. This seems to be more or less the same sense implied in the XML specification, which defines "XML Document" as a string that matches certain formal constraints [xml08]. It also corresponds closely to the concept of an "expression" in the "Functional Requirements for Bibliographic Records" (FRBR) [frbr98] and to some uses of the word "text" in textual philology [tanselle89]. A document in this sense is not an individual physical object[buckland97] but the symbolic structure that can be found simultaneously instantiated in multiple physical objects [derose90].

We motivated the claim that documents cannot be modified in several ways: ^[1]

First, we noted that digital documents, like all digital objects, are typically defined in terms of mathematical entities such as strings, tuples, relations, functions, and graphs. All of these entities are in turn defined as kinds of sets and, consequently, all are extensional: a set cannot, strictly speaking, lose or gain members, and therefore similar constraints will apply to those entities that are defined as kinds of sets, such as strings, tuples, relations, functions, and graphs. A string for instance cannot lose or gain a substring, a tuple cannot lose or gain an element, a relation cannot lose or gain a tuple, and so on. These constraints are consequences of all standard axiomatizations of mathematical set theory. Of course we can say that a string S was transformed into another string S', but this is an idiom -- there is no underlying thing that once was identical with S and now is identical with S', as genuine modification requires. Editing strings is mapping from one string to another, not modifying a persistent underlying entity.

Second, we also observed more generally that abstract objects, and not just mathematical objects or digital objects, appear to have their non-relational properties essentially -- and so have no non-relational properties that they can lose. Since modification requires losing or gaining a non-relational property (ceasing to be thought about by John is not genuine modification) this means that abstract objects generally, and not just mathematical objects or digital objects, cannot undergo modification.

Finally we observed that documents, in the sense intended, are in fact explicitly treated as immutable by communities with fairly practical non-philosophical objectives, such as catalogers [frbr98], textual philologists [tanselle89], and, in at least some contexts, XML specialists.

The Document Modifiability Puzzle

Because the claim that documents cannot be modified contradicts commonly held beliefs to the contrary the arguments just presented can be seen as presenting us with a puzzle, a conflict in our commonsense pre-theoretical beliefs. To systematically explore the various responses to this puzzle, we surveyed the alternative strategies for resolving the following "inconsistent triad":

Responses to an inconsistent triad must reject one (at least) of the assertions in the triad and may be classified according to which assertion is rejected. Each class of responses has a distinctive burden. Responses rejecting the first assertion preserve our intuition that documents can be modified, but must provide an alternative account of what documents are, since they are not strings, graphs, or other immutable abstract entities. Responses rejecting the third assertion typically retain familiar definitions of document as immutable objects, but to be convincing they must provide a plausible account of what apparent document modification really is, since it cannot be genuine modification. Because there appeared to be very little that could be said in defense of rejecting the second assertion (extensional or abstract objects, such as strings, cannot undergo genuine modification) we focused primarily on responses that rejected the first or third assertions.

We found responses that reject the first assertion to be fairly unpersuasive. They required admitting rather challenging ontological entities (e.g. "social objects") and relationships (e.g. "constitution"), and even given these entities and relationships it is not clear how a useful formal account of document change and identity could be developed [cantwellsmith96, [smith03].

This seemed to tip the balance in favor of responses that reject the third assertion: documents can be modified. However a satisfying response here must avoid additional counterintuitive consequences, such as documents that exist before persons, or abstract objects that come into existence outside of the causal framework of natural science. We concluded, adapting the theory of social facts developed by John Searle [searle95], that the least objectionable solution to the puzzle is that a document is indeed an unmodifiable string (or some other similarly extensional or abstract object) in a particular communicative role which it has in virtue of contingent social and linguistic circumstances.

To meet the burden of providing an alternative account of modification we suggested that what is commonly referred to as the modification of a document might be understood as a person or persons coming to prefer one string to another for a particular communication task, and, typically, recording this intention in an information processing system or through some other physical medium. To the extent that our common discourse suggests that there is some thing, some real object, that is genuinely modified, this manner of speaking is not to be taken literally, but as some sort of idiom, metaphor, or "logical construction", perhaps similar to sentences such as "the sun rose in the east," or "the average plumber has 3.2 children". Such sentences can be true, but they cannot be literally true, in the sense of having a compositional semantics and directly (without rephrasing) supporting existential instantiation and valid inference. Of course this cannot be a complete account of apparent change, since it does not give explain "communicative role" in terms that would account for intuitions of apparent sameness in some circumstances of changing preferences, but it nevertheless seemed like a promising direction to explore.

We were under no illusion as to whether we had succeeded in demonstrating that this particular resolution was decisively superior to the others, however we did believe that we had at least accomplished a breadth-first survey of the general kinds of responses that could be made to the triad. So it came as a surprise to realize that we had left one out.

Before turning to a discussion of the missing response we make some brief remarks on the nature and significance of this sort of analysis.

Aside: Why Should Anyone Care?

Other then pure intellectual curiosity, why should anyone be interested in such a strange and abstruse meditation as we presented last year at Balisage, and now correct and extend this year?

We believe that there is a case to be made that reflections of this sort are indeed relevant to the design of information management systems. Inferencing over formal ontologies is increasingly important in information processing, but systematically making information available in logic-based ontology languages requires that the assertions we wish to record be expressed in languages that allow only literal interpretation. Human beings may effectively communicate with natural language sentences such as "The sun rose in the east", "An fog of anxiety descended upon the congregation", "The average plumber has 3.2 children", or "The document was edited". But to support inferencing in the semantic web environment these idioms must be converted to assertions in languages that rely on compositional semantics, existential instantiation, and valid deductive consequences. The underlying ontology need not reflect the latest theories of modern physics, but it should nevertheless at least be internally consistent, and as much as possible avoid clashes with commonsense beliefs. A naive formalization of our familiar discourse about documents does not meet these requirements.

That said, we realize that this sort of thing is not to everyone's taste.

The Missed Possibility: There are No Documents

This spring (2010) at the Graduate School of Library and Information Science, University of Illinois at Urbana-Champaign, we held a seminar titled "Identity and Change in Information Systems", focusing on issues related to formalizing our understanding of digital objects, particularly with respect to representing how they are created and reidentified, and what sort of changes they may undergo. In the course of reviewing prior relevant work on the topic of identity and change we considered a philosophical position, (physical object) eliminativism, which resolves traditional puzzles about change by arguing that ordinary physical objects do not really exist.

Consider the staple problem of "material constitution". A statue S consists of a quantity of bronze B. In unguarded moments we are probably inclined to say that the statue and the quantity (mass, lump, etc.) of bronze are identical. After all they are in the same place at the same time and consist of the same stuff (and not just the same kind of stuff, but the very same stuff, the very same bits of bronze), and appear to have the same causal effects and powers. But problems with this identification ensue.

For one thing if the statue was created this morning from an already existing lump of bronze then the statue has a property that the bronze does not have: being created this morning. And so the statue apparently can't be the very same thing as the lump of bronze. [On the general principle, sometimes known as Leibnit's Law, or the indiscerniblity of identicals, that if x and y are identical then x has all the properties y has.]

Or imagine that our statue S (let's say it is a statue of a horse) is melted and reshaped into another statue, S' (a statue of a person). No bronze was lost so the lump of bronze, B, that was earlier identical with S continues to exist and is now identical with S'. But identity is transitive and so S, the statue of the horse, is identical with S', the statue of a person, -- which seems false.

There are many more puzzles of this sort and many different approaches to resolving them. One rather extreme response is to deny that there are statues and instead to hold that there are just quantities of bronze which have particular properties that they can gain and lose. If the adjective were not already used with a different meaning we could say that some bronze can gain and lose the property of being "statuesque" [chisholm76]. Or alternatively one can say that the bronze can be "arranged statuewise" [merricks01].

The view that things of some familiar kind do not really exist is called "eliminativism" in philosophy. From the eliminativist's point some parts of our commonsense "folk ontology" are inconsistent, mistaken, or redundant, and need to be replaced with a more accurate conceptualization, one in which some of the familiar things in that ontology will have no role. Sometimes the motivation for eliminativism is the view that contemporary science, typically physics, provides a much different and much better account of our world than our folk concepts do. Physicists for instance have routinely claimed that 20th century physics shows that our ordinary concepts of time, space, causality, and physical objects are flawed. One response to these new physical theories is to integrate the findings of physics into our commonsense ontology, which leads to assertions such as: tables are really wave functions. This however is not eliminativism, but reductionism, as it reduces tables to other things but does not eliminate tables from the scientific view of the world -- on the contrary, the existence of the folk object is confirmed not denied in the reduction.

The eliminativist on the other hand is not concerned to provide a more accurate and up-to-date understanding of what tables are, but rather to eliminate tables from our commonsense ontology altogether. The eliminativist claims that with respect to some fragment of our commonsense ontology (in this case medium-sized physical objects) it is not possible to map the concepts in that fragment, concept by concept, into the scientific alternative. Instead the scientific alternative must simply replace, entirely, the relevant portion of our folk ontology. The eliminativist will hold that there is nothing in the real world, as revealed by modern physics, that corresponds closely enough to our concept of table to be appropriately considered a table, and that modern physics will allow us to say all there is to say about how things are in the kitchen without implying that there is a table in the room. Eliminativists argue that the objects of our commonsense notions are not just poorly understood, but that the concepts are too flawed to correspond to anything in the current scientific world view. ^[3]

In philosophy contemporary eliminativism with respect to physical objects is sometimes a response to specific puzzles that reveal a conflict of commonsense intuitions. The problem of material constitution (the statue and the bronze) is an example. Other puzzles motivating recent eliminativism with respect to physical objects include those based on the vagueness of identity conditions for certain objects [chisholm76, [unger79], the unclarity of when it is that two objects are parts of a third thing [vaninwagen90], and the causal redundancy of composite wholes [merricks01]. The eliminativist claims that these puzzles demonstrate the deep incoherence of our ordinary (folk) views and argue that an eliminative resolution is the best recourse available: completely replace our ordinary notions with an alternative system of concepts. Eliminativist arguments are surveyed and criticised in Thomasson [thomasson07]; see also Korman [korman09].

The reader will by now have guessed where this is going: one response to the problem of document modifiability is to simply maintain that there are no documents at all, this would allow us both to avoid accepting obscure social objects into our ontology and also to avoid endorsing the counterintuitive claim that documents cannot be modified. Needless to say these benefits would come at the cost of endorsing another counterintuitive claim: there are no documents.

Why Was Eliminativism not Considered in 2009?

However before going on to explore this alternative resolution we might wonder how it managed to escape the net of last year's systematic survey of possible responses to the inconsistent triad. Recall the triad:

Here is a natural formalization in first order logic:

While the English version of the inconsistent triad may have appeared to be logically inconsistent, this set of first order logic formulas, is in fact not logically inconsistent, at least according to the standard semantics of first order logic. In first order logic a universally quantified conditional "(x)[(isaDocument(x)->isModifiable(x)]" is logically equivalent to "(x)~[isaDocument(x) & ~isModifiable(x)]" and so is true in an interpretation (an assignment of objects to predicates) if and only if nothing in that interpretatin both satisfies the antecedent and fails to satisfy the consequent. This entails that a universally quantified conditional will be true in all interpretations where nothing satisfies the antecedent -- regardless of what the consequent is. An odd and awkward result perhaps (it is one of the so-called "paradoxes" of material implication), but a common feature of first order logic nonetheless. As a consequence of this, in an interpretation where nothing is a document this set of assertions would be all satisfied; and so the set is consistent. ^[4]

Did we actually commit a logical howler in last year's Balisage paper, allowing a class of resolutions strategies to escape our labored survey? Although we did neglect this possible resolution to the document modifiability paradox, we did not, fortunately, claim that a consistent triad was inconsistent.

In our paper we formalized the triad not as in the preceding example, but like this:

And this is indeed an inconsistent set of assertions. The translation of the third assertion as an existentially quantified conjunction was made precisely to ensure that the intuitive inconsistency of the triad expressed in natural language would be present in the formalized version. In the analysis presented last year we focused entirely on solution strategies that rejected the third assertion of this version of the triad only as a consequence of maintaining that its second conjunct could not be satisfied by any individual object that satisfied the first conjunct. Accommodating the possibility that there are no documents was not on the agenda and so we did not consider the possibility that the third assertion is false simply because its first conjunct is not satisfied.

Now that we have identified this overlooked class of solution strategies we propose to remedy the omission by giving this class of responses the same sympathetic consideration we gave the others. ^[5]

In Favor of Eliminativism

If there are no documents our puzzle is resolved. And this is itself perhaps a reason to believe that there are no documents. But it is not much of a reason. To be preferred to other solutions to the document modifiability puzzle the position that there are no documents must have genuine advantages over those other approaches, and perhaps some independent positive rationale as well. Let's see what can be provided along those lines.

We begin by realleging the results of last year's paper. Solutions to the modifiability puzzle that reject the first assertion (documents are strings) and redefine document as a social object, seem to us to be fairly unpromising, for the reasons given before. ^[6] While not decisive, this might nevertheless be seen a providing at least some evidence in favor of the alternative, solutions that reject the third assertion -- though of course this prima facie warrant will be defeated if rejecting the third assertion turns out to have consequences even more obscure or counterintuitive than rejecting the first.

In last year's paper we argued that although not every way of rejecting the third assertion was plausible, there was in fact a coherent position that rejected the third assertion and slightly qualified the first. This approach defined a document as a string in a contingent social role. Documents, being strings, are not modifiable on this account, but document modification discourse could be understood not as expressing literal truths, but rather consisting of harmless idioms and logical fictions that can be paraphrased into more precise representations.

But we now know that there are two ways to reject the third assertion: (i) no modifiable documents, and (ii) no documents at all . So before preferring "no modifiable documents" to "no documents at all" we ought to see what there is to say in favor of "no documents".

Obviously it is commonly believed that documents can be revised, edited, shortened, lengthened, and modified in other ways. This belief is widespread, and deeply rooted. It is perhaps so deeply rooted that it may be an integral part of our concept of what a document is, so intrinsic to the concept that perhaps that nothing can be document if it is not modifiable.

Consider this example from the philosophical literature on the semantics of natural kind terms [kripke81]. We have beliefs about tigers. If some of these beliefs are wrong (say we believe tigers never have more than two kittens at a time, when in fact they often have more than two) then we are mistaken in believing such things about tigers -- but we are still believing something about tigers, even if it is something that isn't true. Now suppose we discover that tigers are not at all what we thought they were, that they aren't mammals for instance, or perhaps not even animals. Suppose they turn out to be robots. Should we really say "tigers are robots"? Intuitions here vary. Suppose it turns out that tiger experiences are being created by sophisticated multi-modal holograms. When we find this out would we say "we now know what tigers really are -- they are holograms"? Or would we say "Well, well, looks like there are no tigers after all, just holograms", or, in a similar vein, "what we thought were tigers are in fact holograms". Of course we might still open with "Tigers are holograms; who knew?", which suggests that there are tigers, and those tigers are holograms. But even if we said that casually, would it really be literally true? Wouldn't it be more accurate to say, after reflection, that there are no tigers, that what we thought were tigers are only holograms.

Perhaps being modifiable is to our concept of a document what being a living organism is to our concept of tiger. If so we could represent this conceptual fact, while maintaining neutrality on the question of whether there are documents, by the conditional: "if there are documents then they are modifiable" ^[7]. But in 2009 we presented arguments to the conclusion that documents are not modifiable. From those two assertions, by modus tollens, the conclusion "There are no documents" follows. The argument, in short, is:

This argument should be taken seriously even if one is skeptical of either (or both) of the two premises. This is because there is some supporting argumentation for each of those premises and, apparently, no decisive arguments against them. The support for the first premise is our argument last year that the unpromising nature of social objects as a way out of the modifiability puzzle provides at least some evidence, though not conclusive, for immutable documents. Supporting the second premise is the widespread and deeply held belief that documents are modifiable. The claim here is not that these considerations are sufficient to justify accepting the conclusion ("There are no documents"), but rather that they seem to shift the burden of proof to anyone who wishes to deny the conclusion and claim that there are documents. Spelled out step by step the reasoning goes:

The most likely response to this line of reasoning will be that the conclusion is on its face more implausible than denying one of the claims on which the argument depends (that modifiability is integral to the concept of a document for instance, or that the social object account of documents should be rejected in favor on immutability).

Making Peace with Eliminativism

The common response to eliminativist arguments against physical objects has been characterized as "Come on!"[wasserman09] Of course there are statues, tables, chairs, documents. What is more obvious? What advantage do we get by the strange tortured refusal to accept the existence ordinary things, especially if it is allowed that those things aren't at all what they seem to be?

The answer to "what do we get?" is presumably: a simple consistent ontology that is free of both puzzles and obscure entities and preserves as many important commonsense intuitions as possible. And this is what we need for safe automatic inferencing in logic-based knowledge representation systems.

But what remains most mystifying about eliminativism (to everyone except eliminativists) is that it will not be satisfied by reducing statues to other things (e.g., "stuff-arranged-statuewise"), but insists that we affirm that there are no statues at all. Why can't it be the case that stuff-arranged-statuewise is a statue? Isn't that what we mean by "statue"? Can't we simply affirm that statues are nothing more than stuff-arranged-statuewise without going on to deny the existence of statues?

The eliminativist rejection of reductionist alternatives is uncompromising. The answer to "isn't stuff-arranged-statuewise a statue?" is: No, there are no statues; there is stuff-arranged-statuewise, but stuff-arranged-statuewise isn't a statue. Of course it is stuff-arranged-statuewise which is the result of sculpting, and stuff-arranged-statuewise which gives art historians statue experiences, but nevertheless: there are no statues. Statues, even reduced to stuff-arranged-statuewise, still create puzzles, and perhaps, in addition, are causally redundant, have inadequate identity conditions, and violate the principles of ontological parsimony... etc. Reductionisms does not address any of these objections.

The eliminativist answer to "but isn't stuff-arranged-statuewise what we mean by 'statue'?" is: maybe, but probably not. Most people seem to mean a thing, a statue, by "statue" and not stuff-arranged-statuewise [merricks01]. Though of course one could mean stuff-arranged-statuewise by "statue"; just as one could mean biscuits-with-gravy by "statue". There should be no argument about that -- the meaning of words is a matter of convention. One's own idiolect can attach any meaning whatsoever to any word. And if "there are statues" is understood to mean "there is stuff arranged statue-wise" then the eliminativist will grant that that sentence is true, and literally true -- but she will still not grant that a sentence understood this way implies that there are statues. This is because the eliminativist does not mean "stuff-arranged-statuewise" by statue when she says "there are no statues. By "statue" the eliminativist means statue, and by "there are no statues" the eliminativist means that there are no statues. ^[9]

But these answers are frustrating if not inflammatory. To get a sense of how one might learn to live without documents, and like it, we begin by taking a closer look at the nature of predication.

When social circumstances put a string in the distinctive communicative role described in last year's paper, why shouldn't one go on to say, as we said last year, that the string is then a document, that this in fact is what we mean by "document", just being a string in a particular kind of role? This is the reductivist complaint to the eliminativist. The eliminativist will answer: because this would have counterintuitive results (documents that are not modifiable), puzzles (things that are modifiable and not modifiable), and violate parsimony (if we can get along without positing documents, shouldn't we?), etc. But such arguments are not decisive when faced with competing claims on the same grounds (what is counterintuitive and what isn't) unless we have a better understanding of what it means to affirm the existence of stuff-arranged-statuewise while resisting statues. Or, in our case here, affirm the existence of strings in roles, but deny the existence of documents.

One answer might go this way. It is often noted that there are many senses of the verb "to be". Identity (represented in symbolic logic as "x=y") is one; predication ("Fx") is another; constitution still another, and there are more. Our business here is with predication. Taking a closer look at predication we can see that it in turn might be divided into two sorts. There is exemplification ("Fido is brown"), and there is instantiation ("Fido is a mammal"). Both of these are represented with the same sort of formulas in symbolic logic (e.g., "Fx"), but arguably they are profoundly different in nature. ^[10]

Making peace with eliminativism might mean simply accepting that "is a document" is like "is brown", or "is a student", indicating that a particular thing exemplifies an attribute, rather than like "is a mammal", which indicates that a thing is an instance of a kind. From "Fido is a mammal" we can conclude that there are mammals, but from "Fido is brown" we do not conclude that there are "browns". Although we can, to be sure, conclude that there are brown things. Of course we can say that there is a document and simply mean by it that there is a thing which exemplifies a role, just like we could (but don't) say that there are "browns" and mean by it that there are things that are brown.

This line of reasoning may seem independent of the preceding argument that there are no documents because if there were documents they would be both modifiable and not modifiable. But there is perhaps a connection. The attraction of the social object approach to documents is that it is consistent with our commonsense view that a document is a mutable instance of a particular kind of thing, a view that supports, psychologically, the development of useful idioms for expressing and recording changes of interest. When we recognize that this perspective is ultimately inconsistent with other things we believe about documents we are perhaps in effect recognizing that document is not a kind with instances; but only a role that immutable strings play in contingent social circumstances.

It is hard to see how to weigh the competing considerations here. If one does proceed to affirm eliminativism on the grounds given then one can tidily resolve the entire document modifiability puzzle without positing counterintuitive unmodifiable documents. But at a cost, of course, of asserting a claim arguably every bit as counterintuitive: that there are no documents. The choice may seem imponderable, and it is not at all clear that restricting existence claims to instances of kinds tips the scale.

Were we already there in 2009?

Last year we observed that the "string-in-a-role" theory of what documents are has as a consequence that on one account of the requirements for being a "type" (or kind) of thing in a formal ontology, documents are not types of things, but rather roles that types of things (strings) have. Guarino and Welty [guarino00] argue that one requirement for a property being a type is that it is "rigid": the things that have that property have it necessarily. The property of being a person, for instance, is a type and meets the rigidity requirement -- a thing which is a person could not have been anything other than a person, and cannot lose the property of being a person. But the property of being a student is not a type, because the things (persons) that are students might not have been students and may, and typically do at some point, cease to be students.

Last year we already had arrived at a sympathetic consideration of the view a document is not a type of thing, but a role that some types of things (e.g. strings) have in particular circumstances. How far is this from eliminativism? One sort of eliminativist, the sort which bases their eliminativism on the kind/attribute distinction just elaborated, might say not far at all. The non-rigid nature of the property of being a document itself indicates that in a sound ontology document is not a type/kind of thing but rather a way that things of a kind (strings) are related. The clash of string-in-role reductivism with our intuitive notion of modifiable documents reidentifiable over time might be taken to reveal that string-in-a-role reductivists have not fully internalized their own analysis, perhaps trying to use the hyphenated phrase "string-in-a-role" to have it both ways, inappropriately reifying a role.

But the string-in-role reductivist will counter that the distinction between kinds and attributes is dubious, and even if real it does not obviously support kind/attribute-based eliminativism with respect to documents. After all, we don't say that there are no students just because we recognize that the property of being a student is not rigid -- being a person enrolled in an educational institution simply is what we mean by "student". ^[11] As to modifiability being integral to the concept of a document, the string-in-role reductivist can simply say we are often wrong about what is or is not an essential feature of a concept. We once thought that documents could be modified, we now know that they cannot be modified.

Where are we now?

Interesting arguments have emerged, and the burden of proof shifts now and then, but the main questions appear to be open still.

Acknowledgements

We are grateful for conversations with many people, but are particularly indebted to David Dubin and Dan Korman. Thanks also to Simone Sacchi and Richard Urban and the other members of LIS590ICI: Identity and Change in Information Systems.