Epistolary Voices. The Case of Elisabeth Wolff and Agatha Deken
Karina van Dalen-Oskam
Huygens Institute for the History of the Netherlands (KNAW)
P.O.-box 90754
NL-2509 LT The Hague, The Netherlands
Abstract
The article focuses on two related issues: authorship distinction and the analysis of characters’ voices in fiction. It deals with the case of Elisabeth Wolff and Agatha Deken, two women writers from The Netherlands who collaboratively published several epistolary novels at the end of the Eighteenth Century. First, the task division between the two authors will be analysed based on their usage of words and their frequencies. Next, any stylistic differences between the characters (letter writers) will be dealt with. The focus lies on Wolff’s and Deken’s first joint novel, Sara Burgerhart (1782). As to the authorship,nothing clearly showed a clear task division, which implies that Deken’s and Wolff’s writing styles are very much alike. This confirns findings of other scholars, who found that collaborating authors jointly produce a style that is distinguishable from both authors’ personal styles. As to stylistic differences in the voices of the characters in Sara Burgerhart, it was found that only a couple of the letter writers are clearly distinguishable compared to the main characters in the novel. I experimented with two possible tools to zoom in on the exact differences between those characters, but the methods are still too subjective to my taste. In follow-up research I will look further than words and their frequencies as building stones of literary style.
1. Introduction
A computational analysis of the stylistic voice of characters in fiction has been the subject of several publications. John Burrows was one of the first. In his book Computation into criticism he analyzed the speech of 44 different characters from novels by Jane Austen using the thirty most frequent function words (Burrows 1987). He showed how even this small amount of high frequency words yielded clearly distinctive results for Austen’s different characters, more so than for the characters in novels written by other authors from Austen’s time or from a later time period.
Recently, John Burrows and Hugh Craig applied multivariate analysis to the speech of characters in a corpus of seventeenth-century plays (Burrows and Craig 2012). They showed that characters indeed can be distinguished in this way, but that the characters of one playwright usually cluster together compared to the characters of another playwright. The author-signal is strongest. In the research of Burrows and Craig, stylometric analysis and authorship distinction are nicely intertwined.
The assumption behind this approach to character voices seems to be that it is a natural wish of an author of fiction to make his or her characters speak thus on paper or on stage as one would expect of a character of that gender, age, social background, etc. But is this true? Could this be based on anachronistic expectations, based on our own horizon of experience? Or is it indeed to be seen as a universal characteristic of all kinds of fiction from all time periods and all cultures?
This assumption should be tested on other (historical) genres. In this contribution I want to make a start with this by trying to find out whether the fictional letter writers in the epistolary novels of two famous Dutch women writers show a significant stylistic differentiation. In theory, this genre offers a perfect occasion for an author to give each of the fictional letter writers their own distinctive voice. This case also involves an authorship problem, however, because the first step will be to check whether the two authors divided their work based on the characters in their story, each writing the letters for a different set of characters. In that case, stylistic differences between characters could also be explained by author difference. After ruling this out, using multivariate analysis from non-traditonal authorship attribution (which one could call a kind of ‘distant reading’), I will then try to zoom in on the details of the differences between characters. For this, not many statistical measures exist yet, which means that close reading automatically becomes the main method for this step. Close reading is an important method that will always be used, also in combination with computational tools. This does not mean, however, that we need not look further and serach for new computational methods and tools. Therefore, I will evaluate several possible computational methods that could go beyond close reading.
This paper thus aims to creatively apply existing computational tools in search of an answer to a concrete literary research question, and to establish for which steps of the analysis more or better tools would be helpful. In this paper, my evaluation of the used tools depends on their empirical strength. Tools that lead to observations that cannot easily be verified and repeated by other scholars I consider to be still too subjective. I will conclude that at this moment, there are no very convincing computational methods yet. This is an example of the still existing gap between ‘close’ and ‘distant’ reading, which I expect will be one of the main topics of research and tool developement of computational literary scholars in the coming years.
2. The authorship case
Elisabeth Wolff-Bekker (1738-1804) and Agatha Deken (1741-1804) met in 1776 and immediately became fast friends.[1] When Wolff's husband died in 1777, Agatha moved in with Elisabeth and from then on, they closely collaborated on many publications, most important of which are their epistolary novels (Buijnsters 1984). The first one, The history of Sara Burgerhart, was published in 1782 and was an immediate bestseller. It was followed by The history of Willem Leevend, a much larger and more complex epistolary novel, published in 1784-1785. The last one was The history of Cornelia Wildschut, almost as long and certainly as complex as their Willem Leevend and published in 1793-1796. Much has been written about the two women and their work. Wolff is known to have been a highly educated and very smart and lively woman, whereas Deken, raised in an orphanage, is described as timid and dull. Based on these impressions, many readers and scholars assume that Wolff was responsible for the lively and funny letters (and/or letter writers), and Deken for the dull and simple letters (and/or letter writers). Even during their lifetime, this seems to have been the general idea. In their forewords and in some personal letters they explicitly stated that this was ridiculous: they did everything in close collaboration. Near the end of her life, Deken states she would like to draw up a list of the fictional letters she wrote, to prove these naïve assumptions wrong. She never got around to doing that. Is it possible, by using stylometric methods, to establish how they distributed the work load between them?
The three epistolary novels mentioned above are digitally available at www.dbnl.org, the digital library of Dutch Literature. They will be compared to three other epistolary novels that are available in digital form in the same digital library: Het land, in brieven (‘The countryside, in letters’) written by Elisabeth Maria Post and published in 1792;[2] Charakters en lotgevallen van Adelson, Héloïse en Elius (‘The characters and tribulations of Adelson, Héloïse, and Elius’) by Anna Catherina van Streek-Brinkman (1804);[3] and De kleine pligten (‘The Smaller Virtues’) by Margaretha Jacoba de Neufville (1824-1827).[4] Finally, the fictional letters will be compared with a digitally available corpus of personal letters written by Deken and Wolff, based on the editions of Dyserinck (1904) and Buijnsters (1987).
3. Authorship results
To find out whether it is possible to distinguish Deken and Wolff from each other in their joint epistolary novels, I start with an overview, comparing all six epistolary novels with each other. I made use of the stylometric R-script developed by Eder, Kestemont and Rybicki, which performs Principal Components Analysis, Cluster Analysis, Multidimensional Scaling, and Bootstrap Consensus Trees (Eder, Kestemont, and Rybicki 2013). I used the last of these for my analysis since the bootstrap consensus tree is a harmonisation of as many different cluster analyses based on word frequencies as the scholar indicates. In each case, I have made several bootstrap consensus trees using ever larger amounts of most frequent words and different incremental steps. Those trees analysing up to 1,000 most frequent words were more stable in form than those only using up to 100 or 500 most frequent words.
The six novels are diverse in length, with a maximum of 585,664 tokens and a minumum of 59,752. In Fig. 1 they have been analyzed in samples of 25,000 tokens. The bootstrap consensus tree is based on a cluster analysis of the ten most frequent words, then fifteen, twenty, twenty-five and thirty most frequent words, so with an increment of five. The further away a sample is from the core of the tree, the bigger the stylistic differences are compared to all the other samples.
Fig. 1: Epistolary novels, bootstrap consensus tree based on the thirty most frequent words
The three Wolff and Deken novels are clearly separated from the novels by the three other (single) authors. This shows that co-authors Wolff and Deken have a distinctive style compared to the other three. A comparable picture arises from the analysis of the text of the thirty main characters from all six novels, all having more than 17,000 tokens in their individual corpus. The characters from novels written by different authors cluster together, and this confirms what Burrows and Craig (2012) showed for British play-wrights.
Fig. 2: Senders from different epistolary novels, bootstrap consensus tree based on the thousand most frequent words
In the next step I will zoom in on one of the Wolff and Deken epistolary novels to find out if the letter writers can be distinguished, and additionally, if the letter writers clearly fall into two different clusters which could be linked to the two different authors. This will be done on their first joint publication, Sara Burgerhart, published six years after they met. This epistolary novel has letters and some other texts presumably written by twenty-six different characters, including one Anonymous and including text written by the authors/narrators. Fifteen of these have a corpus of tokens higher than 2,000 (including the authors/narrators). Although I would rather have used larger samples – as Eder (2013) has shown, 5,000 tokens per sample would yield the most trustworthy results – I prefer to include more than just a couple of characters to make the analysis interesting enough for further exploration. Fig. 3 shows that the different characters are indeed distinguishable when their complete corpora are analyzed and compared.
Fig. 3: Senders from Sara Burgerhart (no sampling), based on the thousand most frequent words
Since the authors/narrators’ text (Forewords and Afterwords) in Wolff and Deken’s novels is usually undersigned by Wolff only, this consensus tree may show a certain work distribution between the two women. Wolff’s style may be prevalent in the letters attributed to the main characters Sara Burgerhart, her women friends Aletta and Anna, and her husband-to-be Hendrik. Through extrapolation, Deken then could be responsible for some of the bad characters in the novel, such as Cornelia Slimpslamp, and Zuzanna Hofland. She could also be the ghostwriter of pious Styntje Doorzicht. And she would be very lively and funny in her letters by Abraham Blankaart. But it is too soon to conclude this; it is unsafe to compare samples of different length to each other. When I use sampling (2,000 tokens per sample), the picture is rudely disturbed and no clear distinction can be found (Fig. 4). Many of the samples are directly connected to the root, which means the software could not convincingly cluster them to any other sample. The only significant branch occurs for the samples of the character Abraham Blankaart, which suggests his letters have a clearly individual style. Blankaart’s corpus of letters is large: 16,362 tokens, but this is not unusually large in the novel: main charcter Sara Burgerhart has 46,392 tokens and Hendrik Edeling 17,627, slightly more than Blankaart. Anna Willis has 14,260. Sara, Hendrik and Anna usually cluster (close) to the root. Only some of the smaller character corpora distinguish themselves from the group. The readers of Sara Burgerhart recognize those characters as the pseudo-religious ‘baddies’ on the one hand (Zuzanna Hofland, Cornelia Slimpslamp) and the overly and therefore sometimes unpractically religious characters on the other (Styntje Doorzicht and Jan Edeling).
Fig. 4: Senders from Sara Burgerhart (sampling), based on the thousand most frequent words
The implication as to authorship is that Wolff and Deken either both worked on the same letters, revising each other’s work all along, or that their style of writing is so much alike that they cannot be distinguished. The first option seems very plausible, but is difficult to prove; historical evidence is not available to confirm this. The second option can be explored by a comparison of the letters in the epistolary novels to the personal letters of Deken and Wolff from the time period in which they wrote and published Sara Burgerhart (Fig. 5).
Fig. 5: Wolff and Deken and senders from Sara Burgerhart, based on the thousand most frequent words
Fig. 5 has three additional corpora in the comparison, from the years 1776-1782. Eleven letters written by Deken are available, 4,594 tokens in total; thirty-five letters written by Wolff, 25,790 tokens in total; and twelve letters that were written jointly, 5,878 tokens in total. The consensus tree again shows no clear distinction pointing to two clearly different authors. And, again, the letters of Abraham Blankaart and some of the (pseudo-)religious characters show up as a rather distinctive set.
What Fig. 5 does seem to confirm is that Wolff probably was the sole author of the Foreword/Afterword. In this analysis the personal letters of Deken and Wolff cluster together as well, but this need not lead to the conclusion that their personal styles are undistinguishable from each other. More tests are needed in which samples from personal letters of their contemporaries are included. Due to a lack of digitized material in this area it was impossible to test this.