Afew years ago, a group of researchers from the Stanford Literary Lab decided to use topic modelling to extract geographical information from nineteenth-century novels. Though the study was eventually abandoned, it had revealed that London-related topics had become significantly more frequent in the course of the century, and when some of us were later asked to design a crowd-sourcing experiment, we decided to add a further dimension to those early findings, and see whether London place-names could become the cornerstone for an emotional geography of the city.footnote1 In the Atlas of the European Novel, Franco Moretti had already worked on the geography of London, mapping residences in Dickens and crimes in Conan Doyle. But emotions have a more elusive reality than buildings or murders, and only one of the Atlas’s hundred images—a map of foreign ideas in Russian novels—was somewhat comparable to the current project. To further complicate matters, when Moretti had shown that image to Serge Bonin, the historical geographer who was advising him about the Atlas, Bonin had been extremely critical: ideas like ‘materialism’ or ‘equality’ were not ortgebunden, as German geographers would say: they didn’t have that intrinsic connection to a specific place which is the basis of every real map. And if ideas were not mappable, how could emotions be?
Then, we encountered a passage in Philip Fisher’s Vehement Passions:
Even more than the ‘geography of fear’, it was Fisher’s remark on the ‘suddenness’ of this emotion that we found illuminating. What is sudden occurs at a specific moment in time, and hence also at a specific point in space: it is definitely ortgebunden, to return to that notion. And if this is so, then a geography of emotions—their actual distribution over a map—becomes imaginable. A London of fear, joy, anger, hopefulness . . .
In programming the study, we began by identifying all proper names in the corpus via a Named Entity Recognition program, later removing from the list those terms that had nothing to do with London, like foreign toponyms, characters’ names, and the like.footnote3 The results are shown in Figures 2 and 3, below. The 382 London locations that had received at least ten mentions formed the basis of our second corpus: about 15,000 passages which—in a version of the keywords-in-context approach—included a specific place-name at the centre, plus the hundred words that preceded and followed it, as in the case of ‘Regent Street’ (see Figure 1). Taggers were then asked to read the 200-word passage and identify the emotion that best characterized it.
At first, we were hoping to capture a wide spectrum of emotional attitudes; but the lack of agreement among the taggers—as well as among the English graduate students who offered to act as a control group—convinced us to reduce the options to the opposite extremes of fear and happiness.footnote4 As a further constraint, a passage would count as ‘frightening’ or ‘happy’ only if at least half of the taggers had identified it as such; and we re-ran all passages through a ‘sentiment analysis’ program.footnote5 And, eventually, some patterns began to emerge. But before coming to them, we need to sketch out the main material transformations of eighteenth- and nineteenth-century London.