Visualizing Les Misérables

Characters and Characters’ Networks in Les Misérables

Michal P. Ginsburg

In 1994 the computer scientist Donald E. Knuth published his seminal Stanford GraphBase: A Platform for Combinatorial Computing. Among the databases he used for demonstrating various methods of combinatorial computing were the characters in Homer’s Illiad and four nineteenth-century novels, including Hugo’s Les Misérables. Knuth’s database for Les Misérables features 80 characters and represents the plot of the novel by listing the encounters among these characters, chapter by chapter.1 This data has been the basis of several graphs, each using a different algorithm to group the characters into clusters or communities.2

Though Knuth and the computer scientists who used his data were not at all interested in Les Misérables as a literary text, their work can be useful to literary critics: as will become clear—I hope—from what follows, it offers a new and productive way of studying characters in the novel. However, the very fact that the interest of literary scholars is in the interpretation of a novel (rather than in combinatorial computing) means that we look at the data (and the graphs) differently. It is because of this difference in goal and purpose that the original database had to be revised dramatically, graphing could not be left entirely to the algorithm, and visualization had to be considered critically.3

·

• I. Methodological Issues

• II. Revising the Data: A Different Understanding of Les Misérables

• III. The Database

• IV. The Graphs

• V. Graphing Characters in the Entire Novel

• VI. Partial Graphs

• VII. Visualization

·

I. Methodological Issues

The very decision to methodically list characters, chapter by chapter, indicates that the goal is to list all of them. But even a quick look at Knuth’s database would show that while it includes some very obscure characters it omits many others: it lists a Mlle Vaubois, friend of Mlle Gillenormand (herself a rather minor character), but does not list Gavroche’s friend, Navet; it includes the Marquis de Champtercier, an ultra-royalist in Digne (with whom Bishop Myriel is shown in one, short exchange), but omits the anonymous man condemned to death, whom Myriel accompanies to the guillotine. The latter example suggests that the presence or absence of a name may be an important factor in the decision to include or exclude a character. In the case of nameless characters (some of whom appear in Knuth’s database), it seems as though appearing in isolation, rather than in close proximity to others, makes characters more visible or memorable. A clear example is the two chapters’ sequence telling of M. Madeleine’s trip to Arras (I, 7, v and I, 7, vii), where in 18 pages we find 16, mostly nameless, characters, none of whom made it to the original database.4

It seems, therefore, that some notion of character as “a person” or “an individual” (rather than as a textual element contributing to the plot and to thematic content) was applied, consciously or unconsciously, in compiling the data; hence the importance of the name as well as of appearing in isolation (that is, appearing at least formally as an individual) and the correlative omission of characters who do not appear as “individuals.”

A second methodological decision, this time clearly deliberate5, was the use of the distinction between “witnessed” and “inferred” encounters as a criterion for what encounters (and hence also what characters) would be included or excluded. This distinction is between what the analysis of narrative calls “showing” and “telling.” There is no doubt that this difference is rhetorically an important one and when looking at encounters we may want to ask ourselves why in one case Hugo chose to “show” us (i.e. dramatize) an encounter and in another to “tell” it. But since the whole novel is told, and much of it is told in retrospect, relative to another point in the plot, to exclude an encounter because it is told rather than shown is problematic. Thus, to take just one example, it does not make literary sense to exclude Valjean’s sister and her children, who were the reason for his stealing the bread and hence for the whole story, because the part of the novel that deals with their relations is told as a flashback rather than dramatized (and I should note that Knuth’s database itself is not consistent in excluding inferred encounters).6

As a result of these two assumptions, the original database omits many characters and encounters and therefore had to be revised, if it were to be used for literary analysis.

II. Revising the Data: A Different Understanding of Les Misérables

In revising the database, small errors in the data (which all graphs ignored and reproduced) were corrected:  for example, Jondrette, father of Gavroche, is not actually a character but merely one of Thénardier’s aliases; the name of Thénardier’s younger daughter is Azelma and not Anzelma. The way many of the nameless characters were referred to in graphs based on the original data was also problematic: “Woman 1” and “Woman 2” does not make for easy recognition. Throughout the revised database nameless characters are referred to in the same way the novel itself refers to them.

But, mostly, the goal of revising the data was to include all the characters that appear in the novel, no matter how insignificant they seem, and all encounters, whether they are shown or told. The result was the discovery that the novel has many, many more characters than readers (myself included) remember or even notice while reading. Most of these forgotten, unrecognized characters are nameless, play a marginal role in the novel’s plot, appear only briefly before disappearing without leaving a trace; they therefore may not be considered to be “characters” at all. I would argue, however, that their presence is of the utmost importance since they stand precisely for “les misérables” of the novel’s title. Thus our habitual reading practices demonstrate the problem Hugo sought to bring to our attention: the invisibility of the miserable ones to the social world we, the readers, represent. Or, put in other words: Realism, which constructs literary characters as “persons” and “individuals,” is complicit with the social order that renders the miserable ones invisible. Recognizing these forgotten, marginal characters means countering our tendency (to which filmic adaptations of the novel contribute) to think of the novel as the story of a small number of unforgettable individuals, rather than as the story of “les misérables.”

The discovery that the novel has many, many more characters than we remember or even notice while reading is well worth reflecting upon. The process of reading is always that of sorting out information, separating what seems to us important from what seems to be less important or not important at all. However, this sorting out depends, at least in part, on cultural assumptions, which it is the purpose of critical reading to reveal and question. Listing each and every character mentioned, no matter how trivial that character appears, no matter how many of them there are, goes against the grain of normal reading and is therefore not a simple task. But the commitment to consider every character and every encounter as equally important—in the sense of equally deserving to be noted—has the advantage of freeing us (temporarily at least) from these presuppositions and allowing us to see all there is in the text.  Since Hugo’s novel deliberately challenges our assumptions and defies our expectations as novel readers, especially by including a huge amount of material that seems irrelevant (especially but not exclusively in the so called “digressions”) it is not surprising that the gap between what we read and what we find in the text is so huge.

The decision to list all characters and do away with the distinction between “real” and “inferred” encounters changed the database dramatically: It now lists more than twice the number of characters—181 instead of 80—and about 500 encounters instead of 77. Doing away with the distinction between “witnessed” and “inferred” encounters eliminated one difficulty in defining encounters but left many others: for example, when Marius spies on the ambush at the Thénardiers, is he part of the encounter? When Eponine enters Marius’s room while he hides under the bed, when Javert recognizes the body of the dead Eponine—are these encounters? The answer to these and similar questions is not cut and dry7 and so an element of arbitrariness, and hence possible inconsistency in listing encounters, remains. And of course, it is always possible that there are still some characters missing.8

Although it may not be flawless, the new database shows us important things about the novel: thematically, as I have suggested, it forces us to acknowledge the existence of the invisible, easily forgettable, insignificant, miserable ones: 93 of the 181 characters—over a half—appear only once in the novel, that is, in only one chapter; 70 of these 93 interact with only one other character. Formally, it highlights the extent to which the plot itself (and not only the digressions that surround it) is dilatory rather than teleological. A good example is the chapter (IV, 8, iv), telling about the complicated plotting of the Patron-Minette criminals, imprisoned in La Force prison, and their friends outside, to rob the house inhabited by Valjean and Cosette—a complicated plotting, involving a large number of characters, that leads to nothing: “rien à faire” (II, 356; “There is nothing to do here,” 1016).9

III. The Database

In revising the database I followed Knuth’s method of identifying each character by a two letter key or code and listing characters and encounters chapter by chapter; like him I did not include historical characters unless they encountered at least one of the novel’s fictive characters so that the digressions are for the most part left out; chapters that are exclusively discursive were also left out. As a result, data is extracted from 285 of the novel’s 365 chapters.

The revised, expanded database has been turned into two interconnected tables, one ordered alphabetically by character code, the other numerically by chapter; one can move from one table to another and thus sort information according to different criteria: name of character, number of chapters in which a character appears, the name and number of characters with whom he/she interacts. Besides providing a convenient tool for retrieving information, it can be used to analyze the relative importance of characters in the novel according to two different criteria: the frequency with which a character appears in the novel and the number of characters with whom this character interacts. For the most part we can expect these two ways of looking at the characters’ importance to correlate: the more often a character appears, the more chance there is he/she will interact with other characters (and hence possibly with more characters). And indeed both Jean Valjean and Marius are the top characters on both these counts. But of greater interest are characters in whose case the two indicators of importance do not correlate. Thus, in terms of number of appearances, the third most important character in the novel is Cosette (who appears in 80 chapters), but she interacts with only 25 characters while Gavroche, for example, who appears in only 27 chapters, interacts with 33. Moreover, while Cosette appears in 80 chapters, her presence in the novel is in fact far greater: as the universal object of desire she is also present/absent in many more chapters where Fantine, Thénardier, Marius and Jean Valjean all think about her in her absence. This makes the small number of characters with whom she interacts even more out of proportion to the other indicator of her importance.

Unlike the graphs, which represent only encounters, the database lists also the appearances of characters alone. These solitary appearances depend, formally, on Hugo’s practice of introducing characters to us before he shows them engaged in the plot, interacting with other characters. This is true not only in the case of major characters; a whole paragraph (in I, 7, i) is devoted to introducing to us Soeur Perpétue who has no role in the plot. Solitary appearances are also the equivalent of monologues, showing characters thinking, dreaming, feeling. And here, not surprisingly, there is a clear difference between central and marginal characters.  That marginal characters are not shown in scenes of introspection is obviously part of their definition as marginal characters; conversely, however, we can say that it contributes to the characterization of the miserable ones by showing them as lacking interiority. Situated as they are on the verge of physical extinction, they do not have the leisure of thinking, dreaming, feeling.

IV. The Graphs

As mentioned above, the data compiled by Knuth was used to generate graphs where the vertices are characters and the edges represent encounters between these characters. It received several graphical forms that differ both in the algorithm used and in the mode of visualization. The graph we produced is different from all previous ones in that it uses the updated data of 181 characters and their encounters.

Different algorithms produce different grouping of characters into clusters; this is because they each have their set of assumptions, which leads to different interpretations of the data. In choosing an algorithm we were looking, first, for one that was articulated enough to be helpful (did not lump all 181 characters in a small number of clusters). We were also looking for an algorithm that would require the smallest number of manual adjustments we felt were necessary to make interpretive sense.10

One reason for manual changes is that an algorithm cannot factor in the relative importance of encounters in terms of their role in the plot and therefore cannot judge properly, in some cases, to what cluster a character should belong. Thus the character Bamatabois appears in the novel twice: he is the dandy who puts snow into Fantine’s dress as she walks the streets of Montreuil-sur-Mer, which brings about the scuffle between them, her arrest, Valjean’s encounter with her, his promise to rescue Cosette—that is, much of the novel’s plot. Hugo also decided, for reasons that are not clear (at least to me) to place Bamatabois on the jury in the trial of Champmathieu, where M. Madeleine denounces himself as Jean Valjean; his presence on the jury, while curious and intriguing, has absolutely no effect on what happens in this scene. But all the algorithms we ran chose to place him in a cluster that consists of all the characters that appear in the trial. From the point of view of his role in the plot this does not make sense and we therefore moved him manually to Fantine’s cluster.

There are other characters who appear only twice and therefore are hard to place. Napoleon encounters both Bishop Myriel and Colonel Pontmercy. To which group does he belong? Here different algorithms gave different answers. In the Girvan-Newman algorithm we ended up using for the whole novel Napoleon is in the cluster around Bishop Myriel, which seems to me the right choice primarily because this is the very first encounter in the novel; putting Napoleon in Myriel’s cluster thus calls our attention to the rather astonishing fact that the very first encounter in the novel involves Napoleon.

Other manual changes involved integrating some 2-character clusters produced by the algorithm into other clusters. It seemed sensible to posit as a minimal number for a community three characters; in addition, some of these 2-character clusters (see for example Maître Scaufflaire and his wife, in I, 7, ii) involve one totally passive character, who is there only to allow the other character to externalize his thoughts. It is important to remember that while our basic premise is that the encounters among characters produce the plot of the novel, characters in a novel (unlike people in real life) sometimes serve a formal need rather than play a role in the plot—which, I would argue, is the case with the passive characters in some 2-character “false” clusters.

Finally, we made a few manual changes to correct errors derived from not taking edge weight (i.e. number of encounters) into account. For example, the unweighted Newman-Girvan graph for the whole novel placed Toussaint in a cluster with Javert, whom Toussaint encounters only once while she encounters Cosette and Valjean (individually or together) 8 times.  Whatever the logic of the algorithm is, putting her in the group with Javert seems to distort the data. Other algorithms placed her either with Valjean or with Cosette and so we manually moved her to Valjean’s cluster.

V. Graphing Characters in the Entire Novel

Increasing the number of characters from 80 to 181 has not increased the number of clusters proportionately. What has increased is the size of some clusters, especially the one around Jean Valjean. As a result we now see very clearly how large the periphery is in proportion to the center: In Valjean’s cluster, the majority of the characters are marginal ones. This modifies our understanding of Valjean’s “centrality” since most of his interactions happen in the periphery of the plot.

The graph also calls attention to the problematic status of his “community,” which is, in fact, a non-community. We normally define community as a group of people whose links to each other are stronger than the ones they entertain with people outside the group. This definition holds true for the group of revolutionaries, for the group of criminals Patron-Minette, and for the bourgeois household of Gillenormand. But it does not hold true for the cluster around Valjean where the characters have a few, if any, ties among them. This highlights another characteristic of the miserable ones: they do not bond with each other; they are marginal because they do not bond and they do not bond because they are marginal.

Inner ties are for the most part absent also in the cluster around Myriel; but here the reason is Hugo’s choice to represent the Bishop through his relations with a large number of very different characters. Again and again we see that formal considerations are as important as thematic ones in determining clusters.

VI. Partial Graphs

The characters’ graph inevitably gives a static picture of the novel’s world of characters. To mitigate this we produced 5 graphs, one for each of the novel’s five parts, using this time the Spin Glass algorithm. The partial graphs show how the novel’s world changes over time. Each of these partial graphs is also a static image, though on a smaller scale. One should, logically, continue by creating a graph for each of the novel’s 48 books; and then, for each of its chapters—at which point the graphs will catch up with and duplicate the process of reading, losing much of the their function.

One of the surprising things that the partial graphs allow us to see is the dramatic difference between Part I and Part II. Part I has the largest number of characters—96—over half of the characters for the entire novel. Part II, by contrast has the smallest number of characters—21—a quarter of the number of characters in Part I; it also has the smallest number of encounters, both absolutely and proportionate to the number of characters. The radical shrinking of the novel’s universe of characters in Part II is partially due to the disappearance of Myriel and Fantine, and with them, of their clusters. But what we don’t necessarily register when reading the novel is that while Jean Valjean does not disappear from the novel, his cluster does: of the 49 characters in his cluster in Part I (this includes the trial cluster which is in fact a sub-group) only 2 remain in Part II (Fauchelevent and Javert) and only very few new ones enter the plot. What we see is that once Valjean rescues Cosette from the Thénardiers, his world shrinks dramatically and never grows back, which explains why when he loses her there is nothing left for him.

In Part III 36 of the 43 characters are new ones, showing the shift of the plot to the Revolutionaries, Marius’s father, and the Gillenormand household. But here the partial graph is somewhat misleading, when it shows the group around Valjean to be the smallest, and hence, implicitly, the least important. In the general architecture of the novel, by contrast, the conflict of Valjean and Cosette vs Thénardier and Javert, receives its most intense treatment in Part III—more precisely, in book 8 of this part (“Le mauvais pauvre”), the novel’s dead center. Evidently, a partial graph cannot represent the role of a part within the logic of the whole.

Part IV shows the increased presence of the revolutionaries due to the group around Gavroche. This increased presence accounts for the fact that part IV has the largest number of encounters (more than Part I, in spite of the latter larger number of characters—96 as opposed to 57).

Part V shows the approach of the denouement as all the major characters—Valjean, Cosette, Javert, Thénardier, Marius, Eponine, Gillenormand—are now combined in one cluster and the groups that dominated the previous parts, shrink (revolutionaries) or disappear altogether (criminals). The partial graphs also show that while Fantine, Myriel, and Gavroche are at different times the center of a cluster, Cosette never is; she moves from Thénardier’s (Part I) to Valjean (Part II) to Marius (Part III). Her centrality to the novel—she is the universal object of desire—is thus quite problematic. We thus start to suspect that no character in the novel is unambiguously central.

VII. Visualization

Independently of the algorithm chosen there is the question of visualization. Though the graphs produced here do not use frequency of encounters to determine clusters, they do give frequency a visual representation: the size of the dot that stands for a character’s place in the network indicates the relative importance of the character, judged by the number of characters with whom he/she interacts. The graph also shows the frequency with which characters interact: the more frequently two characters meet, the thicker is the edge that connects them.

The visualization of data inevitably produces a certain distortion: in order to keep the nodes from overlapping, for example, we need to stretch some of the edges.  Therefore the length of the edges is not proportionate to their weight, even when we take edge weight into account. And yet, when looking at a graph we tend to attribute meaning to the length of the edge. This is sometimes used to insinuate a meaning that does not emerge from either the algorithm or the data.

The graph by Visone (reproduced by Cellier), for example, presents Jean Valjean as the center of the novel, which, as I have argued, is far from the case: the novel is equally centripetal and centrifugal and all centers are problematic. But though we can avoid deliberate misrepresentations, it seems impossible to produce a visualization that does not lead to our attributing meaning to what is, in fact, meaningless. All we can do is make clear that not everything we see necessarily signifies.

1 For the original database see Appendix.

2 M.E.J. Newman and M. Girvan produced a graph in their article “Finding and Evaluating Community Structure in Networks” published in Physical Review, 69 (2004). Jacques Cellier’s essay, “En quête de communautés” (http://jacquescellier.fr/histoire/site_tdh2/fichiersexemples/detection_communautes.pdf), analyzes graphs produced by Pajek and Visone (using the Louvain algorithm). All are based on the original data by Knuth; they represent, however, 77 of the 80 characters listed by Knuth since 3 of the characters listed did not have any encounters.

3 I thank Professor Donald Knuth for taking the time to discuss with me the original data and Professor Mark Newman for answering questions about his graph. I also thank William Parod who first showed me the Girvan-Newman character graph, Philip Burns for producing a first version of the revised database and graph, and Professor Robin Burke for his interest and advice.

4 If one were to summarize the episode of the trip to Arras one would say: M. Madeleine left Montreuil-sur-mer and arrived in Arras where he attended the trial and denounced himself. Everything that happens between his departure and his arrival and self- denunciation would be only a “filler.” But of course the point of this sequence is not what this summary tells us. The point, rather, is that the voyage was full of delays but that in spite of these delays he ended up making it on time. This tension between delay and arriving on time is the correlative of the tension in Jean Valjean between a desire not to go to Arras at all, not to denounce himself, and his conscience, which drives him to go nevertheless; it is also the correlative of the tension between his feeling free to do as he wishes and his feeling under the control of an external force, as well as between chance and fate. Practically every one of the 16 characters with whom he interacts during this voyage to Arras represents this tension and duality—between going fast, being in a hurry to get there, and being stuck or delayed. Thus, for example, the road mender tells him: “Pas moyen d’aller plus loin” (I, 337; “no way to go on,” 249) and then “Le garçon du cheval vous guidera dans la traverse” (337; “The boy that’ll go with the horse can guide your through the crossroads,” 249 ). The Postilion, guiding him: “Voilà le palonnier cassé” (338; “the whiffletree is broken,” 249), but also: “Nous serons à Arras à huit heures” (338; “We’ll be in Arras at eight,” 250). A resident of Arras he questions: “Si c’est un procès que monsieur veut voir, il est un peu tard” (349; “If Monsieur wants to see a trial, he is rather late,” 258-59); and then, when he sees the lit court: “Ma foi, monsieur, vous arrivez à temps, vous avez du bonheur” (349; “Just look, Monsieur, you’re still in time; you are lucky,” 259). The Lawyer he addresses: “C’est fini, dit l’avocat” (350; “’It’s finished,’ said the lawyer,” 259); and then: “[La lumière] c’est pour l’autre affaire qu’on a commencé il y a à peu près deux heures” (350; [The light is for] the other case, which began about two hours ago,” 260).) What pushes Jean Valjean forward and holds him back, what responds to his conflicted desire to arrive on time and to be late, are human beings, agents of fate and of chance, both helpers and adversaries, without whom the entire sequence would have been impossible.

5 In an email dated 1/20/14 Knuth states: “Encounters are generally those that we witness, not that we infer.”

6 For example: Knuth’s database includes Mme Pontmercy, Marius’s mother, even though we never “witness” any encounter with her; the two chapters in which she appears (III. 2. viii and III. 3. ii) are in the “telling” mode and we can only “infer” that she encountered her sister, Mlle Gillenormand, her husband, or her infant son.

7 One way, maybe the only way, of being totally consistent in registering encounters is to adopt a purely formal criterion: if two characters are named in the same chapter, an encounter has taken place. After all, the only space where novelistic encounters take place is the printed page. This, however, would erase the distinction between encounters stricto sensu and remembering, wishing, dreaming—which may or may not be important, depending on what one is trying to determine.

8 It is partially because of the lingering arbitrariness in defining and delimiting encounters that we have decided not to use frequency of encounters to determine clusters. It is worth noting that in many of the algorithms we considered there was no difference in the grouping between the weighted and non-weighted version.

9 All page references are to the Yves Gohin’s edition of the novel (Gallimard, Folio classique, 1995), in two volumes. References to the French text are followed by page numbers in the Lee Fahnestock and Norman MacAffee translation, based on the C.E. Wilbur translation (Signet Classics, 2013).

10 The algorithms we considered are:

Edge betweenness (Newman-Girvan). M.E.J. Newman and M. Girvan: “Finding and evaluating community structure in networks,” Physical Review E 69, 026113 (2004).

Fast greedy optimization of modularity. A. Clauset, M.E.J. Newman, C. Moore: “Finding community structure in very large networks,” http://www.arxiv.org/abs/cond-mat/0408187

Short random walks. Pascal Pons, Matthieu Latapy: “Computing communities in large networks using random walks,” http://arxiv.org/abs/physics/0512106

Spin-glass model and simulated annealing. J. Reichardt and S. Bornholdt: “Statistical Mechanics of Community Detection, Physical Review E, 74, 016110 (2006), http://arxiv.org/abs/cond-mat/0603718

Multi-level optimization of modularity. Vincent D. Blondel, Jean-Loup Guillaume, Renaud Lambiotte, Etienne Lefebvre: “Fast unfolding of communities in large networks.” J. Stat. Mech. (2008) P10008.

Max the modularity measure over all possible partitions. Ulrik Brandes, Daniel Delling, Marco Gaertler, Robert Gorke, Martin Hoefer, Zoran Nikoloski: “On Modularity Clustering,” IEEE Transactions on Knowledge and Data Engineering 20(2):172-188, 2008.

Leading eigenvector of the community matrix. M.E.J. Newman: ‘Finding community structure using the eigenvectors of matrices,” Physical Review E 74 036104, 2006.