1. Background

My interest in “deep” markup began with a remark by my friend and colleague Harold Short, who told me I was doing it. Subsequently, however, I began to worry that I did not deserve the accolade. Then Wendell Piez helped me to realize that in essential respects my markup was thick, not deep. This paper tells the story of how subsequently I have tried to deserve Harold's praise of depth, and how the attempt has led me away from markup altogether, but certainly into the depths of modelling, with questions about markup as a result.

But first a prior story. For the last many years I have been using markup to model the cohesibility of Ovid's Metamorphoses, a classical Latin anthology of mythological tales stretching from creation to the apotheosis of Caesar. One of the most interesting literary phenomena in the Metamorphoses is personification, i.e. the creation of persons from subhuman entities by attributing human-like characteristics to them, as when the trees gather around Orpheus in response to his music. Personification is but one aspect of Ovid's grand project to destabilize all fixed notions of ontology; another is, of course, metamorphosis, as when Daphne changes into a laurel tree. The thoroughness of this project is manifest in the smallest rhetorical detail, at the microscopic level of the narrative, which is often below the level of conscious attention. Ordinarily we think of personification as referring to fully developed anthropomorphic characters who strut their stuff across the narrative, such as Envy in the Metamorphoses, Reason in the Roman de la Rose or Jealousy in the Faerie Queene. But to notice only such “personification characters”, as James Paxson calls them (1994: 35), is to miss most of what trope is doing for Ovid's project. What matters most for Ovid are the threshold phenomena, the brief tremors in normal ontology that happen in the moments when an entity approaches, sometimes achieves and thereafter loses personification. For the most part, such “personification figures” seldom last long enough to take on anything remotely like human form and thus become characters in the narrative. They are anthropocentric but not necessarily anthropomorphic creations of human language.

Three such instances of a personification figure of the abstract fortuna are given in Figure 1.

When one looks into the rhetorical microscope, it becomes obvious that as Morton Bloomfield suggested in the mid 1960s the study of personification should begin at the roots, on the grammatical level, as a phenomenon created by discernable operations in language.^[2] As far as I know, no study such as he called for was carried out on any substantial amount of literary text during or after the flurry of interest his suggestion provoked. Apparently no one at the time thought to follow Fr Busa's lead in getting mechanical help for coping with the massive amounts of data a study of micro-phenomena tends to entail. But to us, now, Bloomfield was obviously right: his insight is grist to our methodological mill. We have the means that he chose not to take up or did not know about.

My work on personification is part of the much larger project I mentioned earlier, to create the Analytical Onomasticon, a meta-indexing tool to all devices of language in the Metamorphoses that refer to persons. As part of it, I have identified a large number of personification figures (somewhere between 500-600) and encoded them in the Latin text along with all the other phenomena for subsequent extraction and manipulation. This is what Wendell has taught me to consider a “thick” encoding -- ca. 55,000 tags for the 12,000 lines of Latin hexameter, or about 4-5 tags per line. Why it is not a deep encoding I will illustrate in a moment in terms of these personification figures.

During the encoding of personification in the Onomasticon I strove to ensure consistency and so reliability of the markup, and to furnish an explanation of it, by formulating what I was then pleased to call a “grammar” of the phenomenon. Subsequently, however, I have realized that given the context in which the work was done this so-called grammar is much better understood as a phenomenology, and that this phenomenology is best regarded as a preliminary sketch toward a genuine grammar. I am, that is, adapting Noam Chomsky's idea of grammar as “a device of some sort for producing the sentences of the language under analysis” (1957: 11).

The impetus to develop a Chomskian-like grammar out of my phenomenology arises directly from a clear implication of Bloomfield's approach: if personification can be studied grammatically, then we should be able to write a set of rules for the trope, or at least for those aspects of it that are grammatical. In terms of markup, if the linguistic factors responsible for a personification can be identified reliably, then they are what should be marked up, not the result, as I did formerly. The personification should be computed from them. Hence the definition of “deep” markup to which I have been led: that which encodes the primitives of a phenomenon and so expresses its grammar, together with the rules for combining them. Insofar as a markup simply, declaratively encodes the phenomena of interest, it is thick, not deep.

2. Heuristic modelling

How, then, is such a grammar to be constructed so that the depths of a phenomenon such as personification can be studied? The approach I have taken is by heuristic, experimental modelling. It is hard to imagine anything better for the job, especially because the phenomenon is poorly understood. In the remainder of this talk I will first describe the model in general terms, then show how it has been implemented and finally comment on the questions it has raised so far.

Roughly the procedure I have followed is as follows: for each instance I record the observable factors contributing to the personification, assign to each a provisional weight with respect to some arbitrary threshold value and compute the result. I resolve discrepancies between this result and my reading of the text whenever possible by adjusting the weights, altering the phenomenological scheme or changing the formula by which the effect is calculated. In engineering terms, the objective is of course to simulate perfectly the literary phenomenon. In scientific terms, it is to identify the inadequacies of the way in which the phenomenon is currently understood and modelled. If the model is properly conceived and the engineering done well, these inadequacies should lead to genuine research questions for further study of the Metamorphoses and to a much improved idea of computational modelling in the humanities.

It should be obvious that the purpose of the modelling as I have framed it is not to prove or falsify some objective truth about the poem. The modelling may approximate a consensus of learned judgement; it may discipline the reading with greater control of and responsibility to the data; but its scholarly purpose is to raise more fundamental questions. It privileges the human reading with all its messy contingencies, then puts that reading to the test. It derives its benefits centrally from the constraints of the medium in which it is constructed -- in this case, total explicitness and absolute consistency of representation -- and from the ability one gains to work at a lower, more primitive level, with something like the building-blocks of literary-critical perception. 

3. Choice of a modelling environment

How, then, is such modelling best done? So far I have been using the term “markup” broadly to denote the act and product of recording a textual entity in computationally tractable form. In the more usual, specific sense, what markup languages primarily have to offer to my research is proximity to the source text. As Stéfan Sinclair has argued in a forthcoming paper on text-analysis (2003), this is actually quite an important matter. In literary studies the chief peril is, perhaps, what Leonard Forster called the “flight from literature” (1978): the constant temptation to escape from the intractable difficulties posed by imaginative language into simplified, abstract schemata. So having the source data immediately present, as the interface for as well as object of study, numbers among the primary desiderata for a modelling tool. In my case, however, it is equally or more important to be working with a data-structuring tool that allows for a centralized, automatically maintained record of recurrent elements and assignment of weights to them. I need to be able to say, for example, that this verb has this much personifying effect on whatever noun of this class takes it as subject, and I need to say this once only. The text needs to be immediately in view, but I must minimize, indeed try always to eliminate any weighting attached to a specific place in the text because such place-specific weightings identify effects, not causes, and so render the result extra-grammatical. More about signs of extra-grammaticality later.

It was at this juncture that another friend and colleague, John Lavagnino, suggested the relational database as an obvious tool. I was immediately persuaded, especially because relational software is highly developed and well understood. Hence my pragmatic choice to model personification by means of a relational database with a spreadsheet as the front-end for calculations and graphical display. Now I am not arguing that a structured markup language could not be involved, certainly not for any principled reason. But life is short, and “real soon now” unacceptable, however technically superior an envisioned tool may be.

4. The relational model

Allow me quickly to describe my relational model by showing a number of the tables involved.

The set of tables shown in Figure 2 are for recording instances of personification, correlating different occurrences of the same entity and assigning to each entity an ontological type. The ontology, shown in Figure 3, is itself an important part of the model, as it roughly expresses an immediate consequence of how personification seems to work on the microscopic level. If, that is, “personification” is defined as an ontological shift to or toward the human state, as I do, then it follows that different ontological types may be differently affected by any given linguistic factor. Thus, for example, in a poem talking to an animal has a different effect than talking to a tree. What matters is what's unusual for the kind of entity in question. The assigned ontological type, as will appear, allows me to give different weights according to the kind.

Weights range from 0 to an arbitrary maximum of 10, which is the designated personification threshold. They are assigned as “independent variables”, i.e. as if each factor had its effect independently of any other. I will return to this point later.

The effect of associated verbs is recorded in a set of tables shown in Figure 4: “Actions”, when the entity is subject of the verb and “Actions on” when it is the object; for both, the “Verb-to-weight” table assigns weights according to the ontology of the affected entity, as just noted. Participles (not shown here) are treated separately and also linked into the Verb-to-weight table. For all occurrences of verbs the grammatical mood is recorded (indicative, subjunctive or imperative). Whatever weight is derived for an occurrence is then multiplied by the value assigned to the mood of the occurrence so that the total effect is appropriately modified. In other words, I build in the assumption that, for example, a command issued to a river has somewhat less personifying force than a straightforward description of that river performing the action. Similarly a wish that something should happen cannot, I assume, be as strongly personifying as the actual happening.

What I call “local factors” — words and phrases in close, grammatical relationship to the entity in question, other than verbs — are treated as in Figure 5. Note especially the PersonifyingType and the Weight columns. The former allows weight to be assigned according to ontology, as with verbs, the latter for crude expression of intangible or unique contribution of the individual occurrence. As noted earlier, the localized weight denotes an undesirable exception to the grammar and so points to the need for more research. It is primarily a way of marking a part of the grammar that is “under construction”.

Larger contexts for the personification are identified in Figure 6. These behave very much like the local factors and so have much the same structure in the database. The same remarks concerning the grammatical exceptions apply here also. Of special note among these exceptions are the “group” and “parallel” context types, which in the current model have no intrinsic weight, i.e. they are entirely extra-grammatical. I will consider these again later.

So much for the database itself.

To this function two modifying terms have been added: one for narrative status, one for the mythological aptitude of the entity in question. The former is intended to allow for attenuating the total effect when a personification is in reported speech, to see if under those circumstances personifications tend to be read as less certain. The latter is used either to amplify the effect for candidates commonly personified in the mythological literature or to attenuate it for those which as a rule are definitely not. I am currently experimenting with these and have reached no conclusion.

5. Modelling

Results generated from the tables by running a set of queries are automatically ported into an MS Excel spreadsheet, from which a chart is then generated. The spreadsheet and chart automatically reflect changes in the database, so something quite close to genuinely interactive modelling takes place. This interaction is another interface issue of considerable importance and an additional reason for choosing a relational database as a modelling tool. Static models are of course unexceptional, but strictly speaking non-dynamical modelling makes little sense, and modelling is what matters here -- and to the humanities as a whole if research rather than production is the goal.

The chart, shown in Figure 7 with the spreadsheet, visualizes personification as the cumulative result of effects derived from the five broad classes of linguistic factors I identified earlier. Context is shown in red, Local Factors in orange, Verbs for subjects in green, Verbs for objects in yellow, Participles in white. You will recall that the threshold for personification is arbitrarily set at 10, the maximum weight allowed for any individual factor. These factors are, you will recall, the primitives of the model: the weights they carry are not calculated but assigned to express a critical reading. These in effect relocate critical judgement to the microscopic level of personification partly to discipline reading with greater control of and responsibility to the data. But the primary purpose of this relocation is to allow comparison of the calculated result with the high-level, mechanically unfiltered human reading, whose processes we would better understand. That is, we would know how we know what we know.

In the early stages particularly, the comparison is bound to show that individual weightings are wrong, either because they are inconsistent or because a particular factor is consistently over- or undervalued. As these teething problems are resolved, however, comparison works increasingly to problematize the manner in which the individually recorded perceptions are combined. Getting the weights to accord with one's sense of the text and the language is relatively easy. The hard problems are combinatorial. 

On the one hand, the assignment of weights to words is relatively easy because the lexicon -- here the Oxford Latin Dictionary -- yields the benefits of an extensive, solidly reliable and radically simplifying scholarship. It simplifies because it is based on the assumption that words in isolation have finite, discernable meaning, and on its basis nicely structured entries have been constructed for these words. Assigning weights is an even more radical simplification of this simplification. On the other hand, combining the weights aims at the full complexity of unmediated literary experience, most of which is out of mechanical sight.

Early trials of the model, however, immediately suggested the now seemingly obvious truth that the accumulation of effects is nonlinear. Across many examples, the degree of apparent personification does not increase at the same rate as the sum of contributing factors. Rather, the impression one gets is of a levelling off, with a point past which additional weight has little or no effect other than perhaps to maintain the personification. Since numerical quantities are involved, the question of how to model this impression is a matter of finding a mathematical function whose curve has the right characteristics.

A common way to proceed is to use a plausible analogy for which such a function is known. The choice is crucial, since the analogy formulates one's best approximation to what the poorly understood phenomenon actually is. It is thus apt to become a strongly influential if not the dominant way of thinking about one's object of study. Analogies are powerful, consequently dangerous.

An analogy that seems right to me is saturation, i.e. the state or process of being charged or filled to the limit of capacity. Saturation has been studied in many physical, chemical, biological and neurological systems,^[3] in many if not all of which the system not only has a certain capacity or maximum response but also an “impedance” or dampening resistance to further change that increases as the system is pushed. One such system is the capacitor Figure 8, a physical device capable of storing electrical energy. I am most grateful to Patrick Juola for suggesting, among several possibilities, the function that models its saturation. 

This function and a curve generated from it are shown in Figure 9. In the equation,

v(t) = V * (1 - e^(-t/RC))

V is the maximum value to be reached, e is a real-number constant (the “Euler number”, approximate value 2.72), t is time and RC is the impedance. For now the important matter to understand is that this function gives an analogically plausible way of calculating a decreasing accumulation of ever more personifying charge, to a specifiable maximum, against an adjustable impedance, at an adjustable rate. For the chart I have set this maximum at 20, to signify a provisional hypothesis that no amount of personifying force will ever amount to more than twice the threshold. I have set the impedance value  RC such that for weights up to the threshold the calculated effect is as close to linear as possible; this allows me to declare by setting an individual weight or perhaps two that an entity is in fact personified.

6. Results and problems

For the sample of approximately 85 candidates currently in the database the results around the threshold and overall are not unreasonable. This is an insufficient number to test the current model adequately, but three very interesting sorts of problems have already become obvious. These are: (1) personification that takes place over more than one or two lines; (2) the interaction of personifying factors; and of course (3) the grammatical exceptions, as noted earlier. I will consider each of these in order.

6.1. Personifications of long duration

For practical reasons I have begun with the simplifying assumption that personification is essentially atemporal, that it has intensity but no appreciable extent. This is in fact not a bad assumption for very brief personification figures, but it rapidly becomes problematic as personification extends beyond a couple of lines. As duration continues, James Paxson argues, the momentary figure rapidly tends toward the personification character (i.e. a fully realized, anthropomorphic player in the narrative). At what point one becomes the other is, however, an open question: is there a discernible threshold? Turning from the model back to the poetry, it is not difficult to derive a possible series of categories from examples, as in Figure 10, and so a corresponding series of thresholds. Are we, then, dealing with several different thresholds and possibly different rates or kinds of accumulation?

To model the new situation would require additional linguistic and contextual factors and a more discriminating mathematical function. Indeed, perhaps beyond level 2 and certainly beyond 3a the problem rapidly converges on narrative characterization as a whole and so requires a very different model. What is needed, however, would seem to be not so much a way of accounting for the number of poetic lines over which a personification is distributed, but rather a way of modelling the distribution. One can imagine, for example, a visualization showing discreet and cumulative intensities plotted against the lines of poetry. I trust Stéfan would approve.

Duration causes at least one other serious and interesting problem to surface. Not only a different kind of saturation is involved and variations in the rate of the personifying “charge” but also the temporal effects of individual factors, which I have so far been able to ignore. The attribution of fatherhood to a river, for example, is not only sufficiently intense in itself to personify but also does so with longer-lasting result than, say, the equally intense attribution of sensibility to a stone. Are personifications like batteries or lovers, some of which last longer? The model needs to show kinds of duration, but it is not at all clear how that is to be measured and modelled.

6.2. Interaction of personifying factors

Another simplifying assumption, noted earlier, is that all the contributory factors are construed as “independent variables”. In some cases, however, attempting to construe them independently illuminates significant interaction. When, for example, the tellus dura -- the earth (tellus) whose dura we may translate either physically “hard” or humanly “cruel” -- throws back the discus that kills young Hyacinthus, the verb, subiecit, may personify tellus itself, but it also would seem partially to personify in collusion with dura, whose anthropocentric meaning it selects. The question this raises is whether in general words have an independently determinable effect of any sort. To return to my example: if I give dura a weight of 5 (out of 10) because half the time it means “cruel”, then in the current model this weight also comes into play when someone falls on the physically “hard ground”. Is there, or is there not, an intimation of cruelty in that hardness? Where does meaning live in a text?

6.3. Grammatical exceptions

I have noted that special pleading, in the form of weights attached to specific occurrences, marks a failure of the grammar and so an area for further research as well as better engineering. The modifying weights found in the LocalFactors table, for example Figure 11, suggest very strongly that accommodating these factors only through the “personifying type” does not reliably account for the effects on various kinds of entities. So, each personifying type appears to need a weighting according to ontology, just like the verbs. More radically, several subtypes need to be identified for the so-called “parallel” and “group” types in the Context table, which are only used to identify two broad classes for location-specific weighting. Here too ontological dependency may need to be figured in.

7. Conclusions

And so it goes: results from the model fail in some particular to correspond to one's sense of the data, temporary adjustments are made, these force an elaboration or reconstruction of the model, new results are generated. The immediate measure of the work is by the better questions one has to ask, in this case about personification in the context of the Metamorphoses. The literary question with which I began was, how in detail personification in the Met might be understood grammatically. I trust you are persuaded that from this question, by an exercise in modelling a provisional grammar of it, a number of better, more precise questions have arisen. I have argued elsewhere that modelling is the best, most comprehensive and fruitful way of talking about what we all do with computers (McCarty 2003), for the simple reason that as Brian Cantwell Smith has said, the computer as a research instrument is essentially a modelling machine (Fetzer 1999). This paper is more broadly an attempt to get at what we essentially do from a concrete instance of it.

But what about deep markup, properly so called? As I noted at the beginning, the current model, built with a relational database manager and a spreadsheet program, lacks the essential engagement directly, visually, on screen, with the text of the poem. An interface for markup, forcing at least the mechanics of this engagement, would address the problem. Use of markup would offer in addition a way of denoting presence within textual structures, which surely is a requirement of a study extending from momentary figures to developed characters. It is also already required if I am to take account of the effects of quoted speech, more so if I extend this to intercalated structures of stories within stories. Does intercalation have an effect on personification? This remains to be probed.

As far as I know, current markup software is, however, non-dynamical. One is able to make models with markup systems but not to do modelling, not properly so called. Identical tag-types are not compiled together in real-time, changes in tagging not immediately reflected in a graphical representation, say. Are there deeper, more consequential impediments, or am I only balking at issues of implementation? I am waiting with some trepidation to hear someone say, “But markup isn't for that. It's only for building permanent, reliably interchangeable structures of knowledge”. If my feared interlocutor speaks for the majority, then I say we have taken a decided turn for the worse and recommend that an emergency council of the philosophically minded be called. Perhaps we're wiser than that, however.

Different data structures should in principle cause one to think differently, just as (but perhaps less obviously than) different interfaces do. I can sense the need for a different interface than I have. Database tables are, I think, not the right way to work with a poem. But what I don't know, the question that I end with, is if, and if so, precisely how, structured markup systems are any better.

As the Psalmist wrote, “Deep calleth unto deep at the noise of thy waterspouts” (Ps 42:7). Depths, when we can hear them calling, are to be explored. But how?