Counting Culture; or, How to Read Victorian Newspapers from a Distance

Abstract

Cultural historians do not like graphs. The linguistic turn and the advent of postmodernism have made us much more comfortable with the ambiguity, plurality and subjectivity of ‘texts’, but correspondingly suspicious about the rigidity of numbers. Rather than search for material or (dare one say it?) objective truths backed by hard data, we prefer to track the emergence of ideas, analyse representations, deconstruct discourse and develop nuanced interpretations of past cultures that can rarely be expressed in quantitative terms. This approach has served us well. Qualitative readings of Victorian culture have provided countless new insights into the ideas that shaped the period. It would be difficult to argue, for example, that Judith Walkowitz's seminal study of sexuality in fin-de-siècle London would have been enriched by a pie chart.1However, it is important to recognize that methodologies based on the close reading of Victorian culture have an important limitation. Put simply, there is too much of it to read. Google's extensive (though by no means exhaustive) database currently lists at least 1.6 million books published in nineteenth-century Britain. To this we must add millions of pages of newspapers, magazines, periodicals, pamphlets, playbooks, advertisements and other forms of printed ephemera. Even the most devoted scholars of the period will only read a fraction of this total. Faced with this mountain of print, we have two choices: to continue subjecting tiny fragments of Victorian culture to close reading, or to supplement this approach by exploring a much larger proportion of the archive through ‘distant reading’. The concept of ‘distant reading’ was coined more than a decade ago (2000) by the literary critic Franco Moretti. ‘[If] you want to look beyond the canon’, he argued: close reading will not do it … . At bottom, it's a theological exercise – very solemn treatment of very few texts taken very seriously – whereas what we really need is a little pact with the devil: we know how to read texts, now let's learn how not to read them. Distant reading: where distance … is a condition of knowledge: it allows you to focus on units that are much smaller or much larger than the text: devices, themes, tropes – or genres and systems. And if, between the very small and the very large, the text itself disappears, well, it is one of those cases when one can justifiably say, less is more. If we want to understand the system in its entirety, we must accept losing something.2 Three years later, Moretti published a series of essays in which he demonstrated three forms of distant reading in action.3First, he plotted the annual publishing rate of novels onto graphs and discovered, among other things, that the genre did not experience a single ‘rise’, as is often argued, but went through several cycles of growth and retrenchment. Similar techniques were used to map changes to the hegemonic form of the British novel and to track fluctuations in the percentage of male and female authors. Second, he created maps and diagrams of fictional British villages and used them to explore how rural class struggle, industrialization and state formation altered the narrative structure of the period's novels. Finally, he drew upon evolutionary theory and used ‘tree’ diagrams to argue that the success of Conan Doyle's Holmes stories stemmed from their author's skilful use of clues. While Graphs, Maps, Trees reads more like a manifesto than a work of literary scholarship, Moretti does not argue that distant reading should replace qualitative approaches to literature. ‘[These techniques] don't offer a new reading of Waverley, Black Forest Village Stories, or I Malavoglia’, he concedes, but aim to understand the larger structures within which these have a meaning in the first place: the temporal cycles which determine the coming and going of genres, or the circular patterns of old village culture, or the stylistic branches that delimit the social function of free indirect style.4 However, this moderate position has since been distorted by journalists and critics who have gleefully branded Moretti as a ‘heretic’. ‘If [he] had his way’, claimed the New York Times, ‘literature scholars would stop reading books and start counting, graphing and mapping them instead’.5Moretti, for his part, has made no attempt to downplay his radical reputation – he is, as Jonathan Arac puts it, ‘deliberately scandalous’.6However, despite his efforts to cultivate a subversive persona, quantitative methodologies are now beginning to edge into the academic mainstream. At the time of writing, Graphs, Maps, Trees has received 232 citations on the Google Scholar database – a number that is worth noting not simply for its size, but because it represents an appropriately Morettian way of measuring the diffusion of ideas. Ironically, a qualitative analysis of these citations reveals something even more significant. Rather than subjecting Moretti to a barrage of reactionary criticism, an increasing number of scholars are beginning to embrace the concept of distant reading and develop innovative ways of quantifying and mapping past cultures. While some researchers are conducting these experiments within established academic disciplines, a growing number of historians and literary critics (particularly those in the early stages of their career) are migrating to the ‘Digital Humanities’. A comprehensive survey of this vibrant new field lies beyond the scope of this essay – partly because it is too dynamic and unstable to be captured by the old-fashioned medium of print.7Each year brings the launch of new Digital Humanities research centres, networks, conferences, journals, prizes, research guides, online communities and academic jobs. While previous attempts to integrate computer technology into the humanities have failed to take off, the Digital Humanities have established a firm foothold in the academy and promise to play an increasingly prominent role. Victorianists would be forgiven for not keeping abreast of these developments. One of the most regrettable consequences of the Digital Humanities movement is the way it has allowed its discussions to develop outside traditional academic disciplines. I suspect that relatively few subscribers to the Journal of Victorian Culture will have read the Digital Humanities Quarterly or attended a dedicated Digital Humanities conference. This is a great shame, for scholars of nineteenth-century Britain have much to gain from engaging with digital methodologies. Victorian print culture is no longer subject to copyright protection and, as a result, an enormous quantity has been digitized, converted into a machine-readable format and uploaded to online archives. One of the most exciting areas of digital humanities research focuses on the use of computational tools to analyse these databases. In 2010, for example, Dan Cohen and Fred Gibbs won one of Google's ‘Digital Humanities Research Awards’ for an ambitious project named Reframing the Victorians. They are currently using ‘text-mining’ techniques to examine the titles of 1.6 million books published in Britain during the long nineteenth-century. By plotting and graphing the frequency of words such as ‘progress’, ‘science’, ‘faith’ and ‘modern’ they have been able to visualize some of the key ideas and forces that shaped Victorian culture.8 A team of Harvard scientists have recently given this particular brand of the Digital Humanities the name of ‘culturomics’.9In their study, they text-mined a corpus of 5 million digitized books and quantified the evolution of grammar, the speed at which society forgets its past, the adoption of new technologies, the effects of censorship, and the changing nature of fame. Best of all, this project inspired the creation of Google's Ngram Viewer – a publicly accessible tool for plotting the frequency of words in the Google Books archive. By way of example, Figure 1 shows the changing ways in which British authors chose to personify America between 1850 and 1900. While ‘Brother Jonathan’ began as the more popular choice, its usage gradually declined. ‘Uncle Sam’ increased significantly in the 1850s, overtook its competitor in 1859, dipped slightly during the 1860s and then exploded in popularity in the mid-1870s. While it is possible to read too much into this data, the linguistic pattern it reveals provides intriguing new insights into Victorian Britain's relationship with the USA. The transition from ‘brother’ to ‘uncle’, for example, is suggestive of a perceived increase in America's power and authority. This kind of research functions best when it is used to target more qualitative readings. For example, by examining the usage of ‘Uncle Sam’ in greater detail, I have linked its rise to growing concerns about a coming ‘American future’.10While Ngram Viewer lacks the sophistication of more advanced corpus analysis tools, the speed and ease with which it allows lone researchers to test new hypothesis makes it a powerful device to read Victorian literature from a distance. A search for ‘Uncle Sam’ and ‘Brother Jonathan’ (1850–1900) using Google Ngram Viewer. Though the Digital Humanities are an interdisciplinary venture, quantitative forms of distant reading have been applied almost exclusively to the study and interrogation of literature. However, the periodicity of the nineteenth-century press makes it ideally suited to the quantitative mapping of cultural change. Each issue of a newspaper can be linked to a specific day in history and, in many cases, to a particular area or social group. While it would be misleading to argue that the press simply reflected the opinions and practices of its readers, the commercial pressures of the nineteenth-century newspaper market compelled editors to be sensitive to shifting tastes. A carefully selected sample of 50 newspapers provides us with a daily portrait of the nation. No other form of Victorian print culture allows us to explore the period with such precision. The potential of newspaper culturomics was recently demonstrated by the digital humanist Kalev Leetaru. Using an archive of 100 million articles published over the last 30 years, he was able to track changes in the ‘tone’ of global news coverage and retrospectively ‘predict’ the Arab Spring and the location of Osama Bin Laden.11It is an eye catching study – one that prompts us to consider how similar methodologies might be applied to our own archives. It is unlikely that a supercomputer will unmask the identity of Jack the Ripper, or settle the case of the Titchborne Claimant but, by subjecting a large corpus of the period's newspapers and magazines to quantitative and tonal analysis, we can gain new perspectives on the Victorian frame of mind. Unfortunately, while hundreds of nineteenth-century periodicals have now been digitized, the quality of the data and the restrictive interfaces of online archives currently make culturomics difficult. The accuracy of optical character recognition (OCR) software is (quite literally) hit and miss; some articles are transcribed accurately into machine-readable text while others are converted into a garbled mess. The recently launched British Newspaper Archive (BNA, 2011) has attempted to address this problem by allowing users to correct OCR text, but the prospect of revising more than 4 million pages is a daunting one. Similarly, while Google has encouraged researchers to download their corpus of books and develop bespoke search tools, the vast majority of nineteenth-century newspapers are only accessible via the interfaces of commercial databases. These archives have been designed with close rather than distant forms of reading in mind. We can hardly blame the British Library for targeting the vibrant family history community rather than the world of computational linguistics, but by hiding their archives behind a pay wall they make it difficult for academics to build their own interfaces.12As a result, research which takes seconds using Ngram Viewer currently takes hours using the 19th Century British Library Newspaper Archive. There are three possible responses to this situation. First, as Michaela Mahlberg argues in a previous contribution to the Digital Forum, we could abandon the press and continue to focus our attention on the quantitative study of nineteenth-century literature.13This approach gives us the freedom and flexibility to develop more powerful and nuanced methodologies, but leads to the neglect of an invaluable source. Second, we could build a new archive of Victorian periodicals that facilitates the development of more innovative research methodologies. While this is an admirable long-term goal, creating these kind of resources is difficult, time consuming and extremely expensive. The British Library – one of the few organizations with the resources to scan Victorian newspapers en masse – intends to digitize up to 40 million pages over the next decade. Unless the Library and its commercial partner grant direct access to this new dataset, the only way to explore it will be through the archive's standard interface. This brings us to our final option: to accept the limitations of existingnewspaper archives and adjust our methodologies accordingly. If we want to investigate this ‘vast terra incognita’ of nineteenth-century print culture immediately, rather than in 10 years' time, this compromise is our best option.14 To that end, I have developed a three-step methodology for performing a rudimentary culturomic analysis of Victorian newspaper archives. While it is far from perfect, it might provide researchers with a useful foundation for more sophisticated methods of enquiry. I have focused my attention on the 19th Century British Library Newspapers. It is currently the most widely used database of its kind and, more importantly, has a relatively forgiving interface for quantitative research. However, the process will work (with some minor modifications) when applied to most other online newspaper databases. The first task is to select a sample of newspapers. The British Library database contains 71 papers. These titles were selected in consultation with press historians and, as a result, amount to a fairly balanced sample. It is tempting, therefore, to immediately begin entering keywords. However, it is essential to recognize that the number of papers fluctuates significantly over the course of the century. For example, there are 38 papers available for the year 1850, increasing to 51 for the year 1880. Without taking account of this changing sample size, graphs will almost always end up displaying an upward trajectory. The simplest response to this problem is to restrict searches to papers that appear throughout your period of interest. Unfortunately, this is easier said than done; while many titles are classified as covering the years between 1850 and 1900, many have months', or entire years', worth of content missing. Figure 2shows the database's coverage for 1850–1900. Red squares indicate the presence of missing issues. While it is possible to discard the worst offenders and focus on the ‘sweet spot’ between 1870 and 1895, it is difficult to eliminate these gaps without losing a significant number of papers and distorting the representativeness of the sample. At present, the best solution is to identify the total number of articles published for a particular period and use this data to ‘normalize’ results.15Rather than chart the raw number of ‘hits’, this technique calculates the percentage of articles containing a particular phrase. For example, rather than claim that 25,164 articles published in 1850 featured the word ‘America’, we might argue that, of the 301,544 articles published that year, 8.34% of them contained a reference to ‘America’. This percentage can then be tracked year by year without worrying about the changing number of titles. Missing issues in the 19th Century British Library Newspaper Archive. Grey squares indicate years in which issues of the paper are missing. The quality of culturomic research is heavily dependent on the careful selection of keywords. A good keyword usually has the following properties: Accurately identifies the idea, object, person, or place under analysis. Does not return a significant number of irrelevant hits. Is not susceptible to linguistic fads (unless language change is the object of study) It not susceptible to OCR errors. This is a stage that I wish could be rushed. Performing quantitative research on a database that was never designed to support it can be laborious. It is necessary to run a separate search for each interval; to gather the number of articles referencing America in 1850, repeat the search for 1851, then 1852, then 1853, … . For complex searches, which take the database a long time to compute, this process can be very time-consuming. The best solution is to initiate 10 searches at a time and return to them later to plot the results onto a spread sheet. The results can then be organized and mapped onto graphs. These approaches, for all their problems, can lead to exciting new discoveries. Here is an example. Figure 3shows the results of three searches for articles in which the names of designated countries appeared within 10 words of the term ‘competition’.16It makes interesting reading. In particular, it reveals an enormous spike in articles linking ‘America’ to ‘competition’ during the 1870s and early 1880s. By zooming in on this spike and repeating the search at monthly intervals, we see that an interest in American competition increases markedly during the autumn of each year – around the time of British harvests (Figure 4). Once these significant patterns are identified, we can subject a manageable sample of articles to closer reading. It is rarely possible to see these wider patterns through close reading alone. To get a sense of the bigger picture, therefore, we need to step back and not be afraid to read the archive from a distance. Articles in which America, Germany and France are mentioned within 10 words of ‘competition’. Results displayed by year. Articles in which America was mentioned within 10 words of ‘competition’. Results displayed by month, January 1878 to November 1882.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 62 citations exist upstream.