Bob Denham quotes Frye saying he is not himself interested in turning literary criticism into a science but thinks it will happen one day. I like to play with this idea. In particular, that there might be an intersection of my two disciplines, literary criticism and computer science, has always been tantalizing to me.
It seems that a scientific study of literature has to begin with a cataloguing of conventions, and by conventions I mean the kind that are more or less obvious to anyone familiar with an area of literature and easily communicable to an outsider. Two books I really love are The Encyclopedia of Fantasy by John Clute and John Grant, and The Encyclopedia of Science Fiction by John Clute and Peter Nicholls. Together I think they contain an excellent start and model for this kind of endeavour. Once this data is collected, we have a set of subjects that can be analyzed, compared and whose general structure can be described. This would produce an “anatomy” of conventions. Just to be clear, by “anatomy,” I do not mean what Frye meant by anatomy. I am using the word to refer strictly to the description of the structure of obvious conventions.
A science of literary criticism could be built up from works of literature to conventions and from a catalogue of conventions to a language or theory for speaking about the structure of conventions, and from this understanding of the structure of conventions to theories about higher level conventions such as genre, and from such higher level theories to a general theory of the structure of literature.
So here are the phases that I believe are needed in a scientific study of literature.
Stage Input Output ---------------------------------------------------------------------- first works of literature conventions second conventions catalogue of conventions third catalogue of conventions anatomy fourth anatomy theories of higher level conventions fifth higher level theories general theory
I think the really essential point is at the anatomy stage. How do we begin to talk about conventions in a structured and objective way? Several answers suggest themselves to me. It may be that we can never get past a semi-formal technical language of convention. But what is more interesting for my purpose is a formal encoding of convention which can be manipulated algebraically and computationally. Admittedly a computational literary criticism would not be a social science as Frye envisioned, but a hybrid discipline combining computer science and literary criticism.
Lets use as our example the “Gender Disguise” convention as described in The Encyclopedia of Fantasy. How do we encode information about this convention? As a first step, lets divide the description of the convention into a series of individual propositions.
1. gd occurs in most genres of fiction 2. consists of a character dressed as and impersonating the opposite sex 3. female to male impersonation is most common 4. occurs commonly in Gothic and mystery 5. in fantasy, gd can modulate into shape-shifting and other magical ways of disguising oneself 6. gd can help create narrative suspense 7. gd motivated by some need or desire of the character to deceive someone 8. occasionally the character is not impersonating but is only mistaken for the opposite sex 9. common scenario is to have a male lust after a supposedly male gd'd female 10. male to female gd usually treated humorously 11. in Arabian fantasy, men often disguise their gender in order to enter harems
What I notice is that qualifying words like “can”, “most”, “usually”, “occasionally”, “commonly”, appear frequently. We can revise our list rewording our entries by removing such qualifications, with the understanding that not every proposition holds universally. This makes propositions 3 and 4 redundant.
I also notice that 1 does not describe the convention nor does it provide any specific context. We should assume that if the context of a convention is restricted in some way, that restriction will be explicitly listed among the properties of the convention. So we remove 1 from the list.
Furthermore, we note that 11 is really one possible instance of 7, so we indicate this with the following notation:
We also note that 8 is a mutually exclusive alternative of 2, and so we indicate this with the following notation
2|8. The vertical bar represents the “exclusive or” logical operator. We now have the following revised description:
2. a character dressed as and impersonating the opposite sex 5. in fantasy, gd modulates into shape-shifting and other magical ways of disguising oneself 6. gd creates narrative suspense 7. gd motivated by some need or desire of the character to deceive someone 8. the character is not deliberately impersonating but is only mistaken for the opposite sex 9. a male lusts after a supposedly male gd'd female 10. male to female gd is humourous 11. in arabian fantasy, men gd in order to enter harems 7<11 2|8
Now we note that
2|8 is a necessary condition and all the other conditions are optional, but that probably at least one is necessary for the convention to be said to be truly present. We can represent this in the following way:
The ampersand represents the logical ‘and’ operator, and the double vertical bar reperesents ‘inclusive or.’
How do we represent context? Consider proposition 11. What this proposition means is that in the context of an Arabian fantasy, men can disguise their gender in order to enter harems. Other forms of gd are not automatically precluded, but men using gd to enter harems does suggest an arabian fantasy context. We can thus reword 11 in the following way:
11. men gd in order to enter harems -> Arabian fantasy
Here ‘->’ represents logical implication. This could be split into two separate propositions:
12. men gd in order to enter harems 13. Arabian fantasy
Eleven is really
12->13. We notice that 12 is a description of the gd convention itself, but
12->13 is a separate claim that does not contribute to the definition of gd. So we can encode the information we have so far in the following way:
gd: (2|8)&(5||6||7<12||9||10) 12->13
We notice that 5 is also not about the content of the convention, but a claim that gd and gender shape-shifting are sister conventions, that is, there exists a more general convention of which they are two alternatives. Let’s call that more inclusive convention “gender deception.” We can define it like this:
And so here is the full encoding:
gd: (2|8)&(5||6||7<12||9||10) 12->13 gender_deception: gd||shape-shifting
Once we have this encoding of conventions, how could we use it? I can think of three main uses.
- Descriptively, to represent existing works of literature.
- Generatively, to construct the conventional structure of possible works of literature.
- Analytically, to demonstrate claims about conventions and develop theories about higher order conventions.
We would not want to manually manipulate these encodings; rather, we would want computational applications. Certainly a computer could generate (2) the conventional structure of possible works of literature, and it could also be used analytically for theorem proving (3). But perhaps more interesting than either than these is the possibility of a computer program reading a work of literature and describing (1) its conventional structure automatically. This is essentially a pattern recognition problem, and would likely require us to rethink our encoding of conventions in order to make them more suitable for existing (and now quite sophisticated and reliable) pattern recognition algorithms.
One class of pattern recognition algorithms conceives of patterns as a set of attributes that are either present (true) or not present (false). There is a certain fuzziness built into these algorithms in the sense that, for a given pattern to be detected, not all of its attributes have to be present, and those that are present do not always need to be successfully detected.
To continue with our current example, we can have true/false tests like:
Is there a harem? Are there female clothes on a male body? Is there suspense? Is there a magical alteration of the body? Is there an intent to deceive?
And so on. What we find is that these questions, though straightforward for a human being to answer, are themselves pattern recognition problems in their own right. If we wanted to fully automate the process of finding conventions in literary texts, we would need to discover an algorithm that could answer questions like these with a reasonable amount of certainty. Some of these questions might be answerable by use of text analysis tools, but others would likely require natural language processing.