TEXT IF JOTTINGS PLUS PROCEDURE (JoPP) .dvi, .ps versions available through my web site, under "In Progress Publications." Files are http://www.rpi.edu/~brings/MONISTEL/monistel1.dvi http://www.rpi.edu/~brings/MONISTEL/monistel1.ps http://www.rpi.edu/~brings/MONISTEL/monistel1.txt Selmer Bringsjord Dept. of Philosophy, Psychology & Cognitive Science Department of Computer Science Rensselaer Polytechnic Institute Troy, NY 12180 USA selmer@rpi.edu o http://www.rpi.edu/~brings January 4, 1996 Abstract Allen Renear's excellent target paper describes a number of general po- sitions on the question of what text is. I find none of them satisfying. By my (confessedly idiosyncratic) lights, text is ultimately composed of jottings plus procedures for unpacking these jottings. Such a view marks an attempt to take seriously, via formalisms in use in the field of Artificial Intelligence, an interesting thought-experiment given by Wittgensteinian (which, incidentally, W thought supported the idea that the mind can store more information than the brain). My view seems to be supported by most of the advantages and arguments Re- near cites in favor of the "Text is an OHCO" view, but seems to be a Platonic (in Renear's sense of the term) position unthreatened by the difficulties confronting the Platonic OHCO view. 1 The Fundamental Intuition Because my take on text is bound to baffle, it may help if I share, at the outset, my fundamental intuition_which is that text, at bottom, is informa- tion represented in its purest, more "unadorned" form. The best candidate for such a form, it seems to me, is logic. My intuition is to some extent brought to life by intelligent computer systems. Such systems (e.g., BRU- TUS, featured in [1]) are able to generate a finished text from information represented in austere logical form. 1 2 Wittgenstein's Thought-Experiment 2.1 Wittgenstein imagines someone (J) jotting down inscriptions as someone else (R) recites a text, where the jottings are necessary and sufficient for J to reproduce the document in its entirety. "What I called jottings would not be a *rendering* of the text, not so to speak a translation with another symbolism. The text would not be *stored up* in the jottings" ([7], 612). Wittgenstein goes on to ask: "And why should the text be stored up in our nervous system?" (612).1 This question, given the focus of our Monist dialectic, is tangential. We aren't in the business here of investigating whether mental activity always corresponds to neurophysiological activity in the brain. But the sort of jotting to which Wittgenstein draws our attention here is, as I will explain, suggestive of what I regard text to at bottom be. 2.2 In order to fix the thought-experiment, suppose that J jots down a list L of 5 bullets, o u1 o u2 o u3 o u4 o u5 where each ui is associated with a short string from some natural language. Suppose, in addition, that R recites an essay E of over 2000 words. We assume, as well, that J can, at any point after hearing R's essay, reproduce E from the list L. So far, the gedanken-experiment involves characters, actions and objects interacting in a manner we could certainly witness in the "real world." But let's extend things just a bit: Suppose that J commits L to Footnote 1: Hao Wang [6] recently recounts that in trying to reconstruct discussions he had with Kurt G"odel in the seventies he drew on notes taken at the time. Wang believes that in some sense his brain contains traces of the interaction in question not represennted in the notes. He also takes seriously the possibility that his mind remembers more than the traces in his brain. memory, so that he no longer needs aids like pencil and paper to reproduce E. 2.3 My view is that E, at bottom, is L plus whatever procedure allows for the expansion of L into E. More generally, my view is that text is really, at bottom, jottings plus procedures (for reproducing a final text, where such a text can be in written or oral form). I hereby baptize this view with it's own slogan: Text is JoPP. 2.4 The JoPP thesis requires elaboration, clarification and defense beyond what I offer in this short commentary. This I readily admit. Here I seek merely to point to the view, and to get it on the table. 3 In Defense of JoPP 3.1 Renear cites certain practical publishing advantages of content-based text processing (sect. 4.3). He also cites a list of arguments in favor of the Platonic OHCO thesis, where the first of these arguments is essentially re- capitulation of the practical advantages (sects. 5.1.6-5.1.1.10). Once the JoPP thesis is incarnated via certain models and techniques in use in the field of Artificial Intelligence (customarily abbreviated as `AI'; the field is devoted to engineering intelligent systems such as expert systems and robots), the thesis is perhaps supported by all of Renear's arguments for the OHCO view. Not only that, but the objections to Platonic OHCO which push us toward Pluralism and Antirealism, objections Renear is inclined to resist, fail to threaten the incarnated JoPP position. 3.2 One approach to AI, the so-called *logicist* or *symbolicist* approach, represents the knowledge, belief and reasoning of sophisticated agents (including human agents) in a logic. Often, the logic used is a particularly well-understood one, namely first-order logic (FOL) [5], [2], [4]. For our dialectic it's not necessary that we discuss the technical details of FOL. (Renear himself, as his own AP- 2 indicates, is familiar with FOL. In the Appendix I provide a rapid overview of FOL.) The important thing is to note that in the logicist approach to AI, a document in natural language (say a story involving betrayal, written in English) is "compressed" to a set of formulas. In other words, the document is captured by certain jottings (though here the jottings aren't, as was the case with u1-u5, expressions in natural language). Given certain algorithms, the jottings, or formulas, can be used to reproduce the story. (For a look at some of the algorithms in question, see the chapter on language processing in [5]. For research which employs FOL jottings and these algorithms in order to construct artificial storytelling systems, see [1].) Accordingly, suppose that, in accordance with the JoPP thesis, our essay E from above is captured by a set of formulas from FOL, from which, by some algorithm A, E can be recovered. How does this view of E, and, in the general case, this view of any document, gain support from the arguments Renear cites? Let's consider the arguments in turn. 3.3 Pragmatic/Scientific Consider first the argument (sect. #5.1.6) that since treating texts as OHCOs yields practical advantages (text as an OHCO is easy to create, modify, print, transfer, etc.), texts ought to *be* regarded to be OHCOs. If, as the logicist approach to AI predicts, the future will bring reliable and well-oiled procedures for capturing documents as JoPP, and for going from jottings to fully developed documents, this argument will go through equally well for treating text as JoPP. For now, whether argument one supports the JoPP thesis is at worst an open question. (It may be worth noting that *people* often capitalize on the utility of a JoPP-like approach. While the ability to reproduce documents verbatim, in the manner of our hypotheti- cal J, is rare in our world, people certainly often *approximate* J's behavior. This paper, in fact, was generated from a list of bullets; and even now, I understand what I have written by understanding a small set of jottings.) 3.4 Empirical/Ontological There is little question that this argument (sect. #5.1.7) supports the JoPP thesis. This is so because our talk about texts often makes use of jottings. We categorize entire novels (and entire works of philosophy, poetry, etc.) by single words. (E.g., "Flaubert's *Sentimental Education* marks a high- point in literary realism." Or: "Smith's book, while published in 1994, is fundamentally Cartesian.") We_and here our behavior is in line with the AI incarnation of JoPP_talk about the underlying propositions and arguments advanced by an author, about certain truths present in a given text, and so on. 3.5 Metaphysical/Essentialist This argument for the OHCO view (sect. #5.1.8) is easily adapted so as to support the JoPP thesis. There should be little doubt that if, for example, the typeface of a document is changed, the text is not. But just as changes to the content objects of a text produces a numerically distinct text, so too changes in the FOL distillation of a document produce a different text. On the AI-incarnated JoPP view, two documents are identical if and only if they are generable from the same core information in a certain logic. 3.6 Productive Power Renear claims "that an OHCO representation of a text can mechanically generate other competing representations but none of these other other representations can mechanically generate an OHCO representation" (sect. #5.1.9). This claim, with `OHCO' supplanted with `JoPP', yields a true statement. In fact, perhaps the chief virtue of jottings (whether in the form used by J, or in the logical form used in language processing in AI) is that they have great productive power. 3.7 Conceptual Priority To grasp a text, Renear tells us, is to grasp the OHCO structure of a text. About this I'm afraid I'm somewhat skeptical. It seems to me that to grasp a text is to grasp the propositions underlying that text. To grasp such propositions is to grasp the kernel captured in jottings. (In symbolicist AI, understanding a story consists in arriving at command over a set of formulas from some logic.) But it would seem that one can grasp the underlying propositions of a text without attending to OHCO structure. (Imagine a mathematician grasping the underlying propositions in a published proof without attending to OHCO structure.) 4 JoPP's Relationship to Pluralism and Antirealism 4.1 Pluralism The observations that force modification of OHCO Platonism toward what Renear calls "pluralism realism" (sect. #5.2.10) are ones the JoPP view accommodates from the outset: the JoPP approach is designed to allow for distillation of disparate documents. Whether the final text is a short story, a proof, a physics textbook, a poem, whatever, the JoPP thesis is that such text can be captured as a set of assertions in a formalism like first-order logic. Put another way, the JoPP view is that in order to build an artificial intelligence capable of generating short stories (etc.), one must first ensure that the AI in question has command over the core information_from which, via various algorithms, the final text can be generated. 4.2 Antirealism 4.2.1 It is interesting to note that there is a fundamental clash between the JoPP view (and the related approach to building sophisticated artificial agents) and the first Huitfeldtian [3] argument Renear cites in favor antirealism (sect. #5.3.7). The JoPP thesis entails that there is a key set of facts about a text which are thoroughly objective. 4.2.2 The second rationale in favor of antirealism_that there are many diverse methodological perspectives on a text_is one that the JoPP approach em- braces. In order to produce different kinds of structure (physical, composi- tional, narrative, etc.), the procedures going from jottings to final text need only be suitably adjusted, but the jottings needn't change. 5 Objections, Very Briefly 5.1 Objection 1 5.1.1 It will doubtless be said against me that my JoPP view of text is really a view of something that *underlies* text. As such, my position doesn't address the issue at hand. 5.1.2 In response, it's true that Renear characterizes theorizing about text en- coding as revolving around two fundamental categories, namely "linguistic content" and "additional information" related to this content (cf. #3.0.1 in the target piece). But the phrase "linguistic content," given our emphasis on the theory of text, should be elastic. The title of our dialectic, after all, is "*Philosophy* and Electronic Publishing." 5.2 Objection 2 5.2.1 Some may object that, at present, the JoPP view, insofar as it is wed to AI, is but a pipe dream: there are no artificially intelligent systems around today that can grasp and generate complicated text. 5.2.2 The obvious reply is that in searching for the essence of text we need not be constrained by what is currently feasible. Besides, for some moderately complicated texts, JoPP is already instantiated in some working computer programs. Appendix: Super-Brief Account of FOL (This section is sure to be ugly in the ASCII version of this document. For a pretty .dvi or .ps presentation of the underlying .tex file, go through my web site.) Given an alphabet (of variables x; y; : :,:constants c1; c2; : :,:n-ary re- lation symbols R; G; : :,:functors f1; f2; : :,:quantifiers 9; 8, and the familiar truth-functional connectives (:; _; ^; !; ) one uses standard formation rules (e.g., if OE and _ are well-formed formulas, then OE ^ _ is a wff as well) to build "atomic" formulas, and then more complicated "molecular" formulas. Sets of these formulas (say ), given certain rules of inference (e.g., modus ponens: from OE and OE ! _ infer to _), can lead to individual formulas (say OE); such a situation is expressed by meta- expressions like ` OE. First-order logic, like all logical systems, includes a semantic side which systematically provides meaning for formulas involved. In first-order logic, formulas are said to be true (or false) on an interpre- tation, often written as I |= OE. (This is often read, "I satisfies, or models, OE.") For example, the formula 8x9yGyx might mean, on the standard in- terpretation R for arithmetic, that for every natural number n, there is a natural number m such that m > n. In this case, the domain of R is N, the natural numbers, and G is the binary relation > N x N, i.e., > is a set of ordered pairs (i; j) where i; j 2 N and i is greater than j. In order to concretize things a bit, consider an expert system designed to play the role of a guidance counselor in advising a high school student about which colleges to apply to (I have such a system under development at Rensselaer). Suppose that we want a rule in such a system which says "If a student has low SATs, and a low GPA, then none of the top twenty-five national universities ought to be applied to by this student." Assume that we have the following interpreted predicates: Sx iff x is a student, Lsx for x has low SATs, Lgx for x has a low GPA, T x for x is a top twenty-five national university, Axy for x ought to apply to y. Then the rule in question, in first-order logic, becomes 8x8y[(Sx ^ Lsx ^ Lgx ^ T y) ! :Axy]: Let's suppose, in addition, that Steve is a student denoted by the constant s in the system, and that he, alas, has low SATs and a low GPA. Assume also that v is a constant denoting Vanderbilt University (which happens to be a top twenty-five national university according the U.S. News' annual rankings). These facts are represented in the system by Ss ^ Lss ^ Lgs and T v: Let's label these three facts, in the order in which they were presented, (1), (2), and (3). Our expert system, based as it is on first-order logic, can verify {(1); (2); (3)} ` :Asv; that is, it can deduce that Steve ought not to apply to Vanderbilt. References [1]Bringsjord, S. & Ferrucci, D. (forthcoming) Artificial Intelligence, Story Generation and Literary Creativity: the State of the Art (Hillsdale, NJ: Lawrence Erlbaum). [2]Genesereth, M.R. & Nilsson, N.J. Logical Foundations of Artificial In- telligence (Los Altos, CA: Morgan Kaufmann). [3]Huitfeldt, C. (1995) "Multi-Dimensional Texts in a One-Dimensional Medium," Computers and the Humanities 28: 235-241. [4]Pollock, J. (1995) Cognitive Carpentry: A Blueprint for How to Build a Person (Cambridge, MA: MIT Press). [5]Russell, S. & Norvig, P. Artificial Intelligence: a Modern Approach (Englewood Cliffs, NJ: Prentice-Hall). [6]Wang, H. (unpublished) "On Computabilism: Some Subproblems." [7]Wittgenstein, L. (1967) Zettel, trans. Anscombe, G.E.M. (Oxford, UK: Basil Blackwell).