Enhancement, Transformation, and Zukunftsmusik: Established methods of interpretation transposed into the computational realm.

One of my favorite literary genres is French twentieth-century crime fiction, roughly from the 1930s (Georges Simenon) via the 1950s (Léo Malet) to the 1970s (Jean-Patrick Manchette). This is quite a large area, so one of the things I would like to do, if the state of text digitization and copyright law will permit it some day, is to do computational text analysis on a large collection of such texts.

This is the cover of the 1970s band Factory’s album “Cache ta joie”, the soundtrack to Jean-Patrick Manchette’s only theater play. Photo credit: Benardo Le Challoux, http://www.flickr.com/photos/bernardolechalloux/4548081612/, CC-BY-SA – http://creativecommons.org/licenses/by-sa/2.0/. (There do not seem to be any CC-licensed images of Manchette.)

For now, what I’m doing is read more crime fiction on the one hand, and read some critical works on it on the other hand. And do so, of course, with this project in mind of doing distant reading instead of massive close reading. Existing critical approaches are usually based on “normal” reading practices of individual authors’ works or of more or less representative selections of crime fiction. Can such approaches be transposed to a quantitative, computationally-supported method of analysis? How do they scale, i.e. how difficult is it to move from minute individual observations and their careful interpretations to exhaustive collection-level query results and their interpretation? And how does the focus of possible questions shift due to such transpositions? (In a way, this post goes back to a set of questions I have adressed earlier, mainly around interpretations of Rousseau.)

To take a concrete example, let’s focus on the book I have been reading these days: it’s an analysis of Jean-Patrick Manchette‘s work, written by journalist and critic Frank Frommer, and published in 2003.[1] This is a solid work of literary criticism without being particularly innovative in the way it constructs its argument, presents individual reading observations or contextualizes them in relation to larger social trends. Two types of interpretive strategies have struck me particularly: one, where short sections of each novel are taken to symbolize and prepare major meaningful structures of each novel as a whole; another, where a specific and recurring, relatively localized stylistic or lexical phenomenon is observed and given a meaning as a pattern.

An example for the first type of interpretive strategies is the way Frommer takes the opening scenes of each novel as a means for Manchette to establish not only a tone for the novel, but also to create a certain interpretive attitude in the reader; also, an abstract structure in the initial scene is taken to be representative of an important semantic or even ideological structure in the novel.[2] For example, when in Fatale, one of the male hunters is suddenly transformed from hunter to victim, and the female protagonist becomes the dominant figure who kills the victim, this comes to be symbolic, for Frommer, of the more general principle of a reversal of established hierarchies (dominated vs. dominating characters) in the novel which therefore becomes a narrative of a critique of bourgeois ideology.

How could such an approach be scaled up to a quantitative approach? How can we pursue this type of interpretations, discover these kinds of structural analogies, when working not with Manchette’s ten novels, but with a collection of hundreds of novels? Maybe something like sentiment analysis applied to the protagonists could be one possible direction. Or maybe go directly to AI for modeling of narrative structures. In any case, these are hard questions without obvious solutions. Because this reading strategy operates not on the level of the lexicon, but on abstract constructs of meaning and symbolic significance of large narrative structures, identifying them computationally is a thing of the future (“Zukunftsmusik”, as we would say poetically, for once, in German).

An example for the second type of interpretive strategies is the way Frommer follows up certain lexical patterns and identifies commonalities in their use in order to give their usage a specific meaning.[3] For example, he notices the recurrent use of the adverb “vaguement” (vaguely) in one of Manchette’s novels. This recurring use is in an intriguing relationship to the “behaviorist” style of narration Manchette is practising in the wake of Hemingway of Hammett (also described as “exterior focalisation” by genettians). Another example is the way he analyses the use of names for characters or of brand names to describe (or rather, instead of describing) cars and many other commodities. Frommer relates this latter phenomenon to the more general theme in Manchette’s work of a marxist alienation of modern man from society and him/herself.

This type of reading strategy, rather than being a challenge to computational approaches, seems clearly to be facilitated, enhanced, and deepended by them. Instead of noticing one specific adverb in one novel, one could search for this adverb and a whole list of its equivalents and near synonyms in the entire collection of novels. Rather than just looking at one or two examples more closely, a “keyword-in-context” visualisation of the results or, in a slightly more advanced way, a collocation search, could yield interesting insights which it would take a long time to come up with using more traditional reading-and-annotation techniques. So this kind of strategy should scale easily, but it also replicates relatively beaten critical tracks. Of course, rather than starting from an initial reading observation or intuition, one could also do topic modeling on the entire collection and may (or may not) discover a “brand name” topic present in certain novels but not in others. The outcomes of topic modeling, however, are really exploratory in nature and are not well suited for very targeted queries aiming to confirm or infirm specific existing observations (at least not in this way).

In any case, what follows for me is that there are three ways of transposing established close-reading methods of interpretation to the computational realm: first, replicate and enhance existing lexical pattern-finding strategies by scaling them up to the collection level; second, approach those same lexical reading strategies with genuinely quantitative methods, something which leads to a transformed reading strategy; and third, marvel at the human capacity to understand, synthesize, abstract and compare large complex narrative and argumentative structures; here, we are probably limited to start by defining what we would need to get even a modest head-start on such problems. Honestly, I think these are three great options and I wouldn’t want to have to choose among them. Would you?

1. Franck Frommer: Jean-Patrick Manchette. Le récit d’un engagement manqué. Paris: Kimé, 2003. There is a review of this book over at Fabula.org. [↩]
2. Frommer, p. 37-49. [↩]
3. Frommer, p. 49-66. [↩]