Sentiment analysis (SA) and emotion detection have become an approachable methodology not only for scholars with a background in information technology and natural language processing (Pang and Lee), but also for those in the humanities (Y.-S. G. Kim et al.; Eder et al.; Scherer; Schindler et al.). A growing body of SA studies has determined sentiment scores for whole literary texts or computed affective variation throughout a narrative (E. Kim et al.; Reagan et al.; Jacobs; Mohammad, Sentiment Analysis of Mail and Books; Klinger et al.; Zehe et al.). Under investigation ar different types of texts (Kakkonen and Kakkonen; Jockers and Underwood) and genres such as fairy tales (Alm and Sproat; Mohammad, “From Once upon a Time to Happily Ever after: Tracking Emotions in Mail and Books”), and historic plays (Mohammad, “From Once upon a Time to Happily Ever after: Tracking Emotions in Mail and Books”; Schmidt et al.). Results are however often not suited for generalisations: on the one hand, because of the lack of standard procedures recognised by the academic community, and on the other, because of the idiosyncratic characteristics of corpora representing particular literary domains, as well as because of the specificity of sentiment and emotions lexicons, often built to address specific research questions.

As for many other complex phenomena, literary studies have imported theories of emotions and sentiment from other disciplines, in particular cognitive psychology, evolutionary biology and philosophy. These differ in their conceptualisation of such phenomena as well as in their application of these, and it is therefore not surprising that several different concepts of sentiments and emotion are used in literary studies.

We conceptualise here textual sentiment as the underlying feeling, attitude or evaluation associated with a proposition that can be analysed in terms of valence and arousal (Liu). The term valence reflects the sentiment orientation, or polarity, ranging from pleasant to neutral to unpleasant. One can think of it as an indication of ‘positivity,’ i.e. whether the general sentiment of the text is negative, neutral or positive. Sentiment can also be evaluated in terms of arousal, which is defined as the degree to which a particular stimulus ‘activates’ the experiencer for action (Russell, “A Circumplex Model of Affect.”). Arousal ranges from high (or active) to low (or passive), and can be substantially understood as the level of intensity of the sentiment expressed.

The concept of emotion is somewhat more complex, and it has been object of numerous debates, especially with regards of what is perceived and defined as ‘emotion’ and what can be empirically observed (Scherer). While the open questions on this topic are certainly worth further research, we will treat ‘emotions’ here as understood by Ortony and Turner and Clore and Ortony, i.e. as ‘affective reactions made specific by the different social and psychological situations in which they arise. […] their identity depends on the context. As a result, emotions can be as varied in nature and number as the situations they represent […] Emotions, then, are simply affective states that are about something. (Fox et al.)’

According to several psychological models of emotions, humans are thought to have a set of so-called ‘basic emotions,’ particular affective states that are assumed to be cross-culturally recognizable. Basic emotions are described as ‘discrete’ because they are believed to be idiosyncratic and distinguishable by an individual’s facial expression and/or biological processes. Such ‘basic’ emotions can combine to form more complex ones to cover the whole emotional spectrum. Some of the existing theoretical models for basic emotions are summarised in Table 1 below:

As readers, it is generally possible to intuitively understand the sentiment and the emotions conveyed by a text, thanks to linguistic devices as well as to co-textual and contextual information, including individual or shared socio-cultural experiences.

What we are interested in examining in this paper is how these values are encoded in the text, rather than how readers respond to them. In order to express emotions and sentiment in writing, several linguistic strategies can be used, such as grammatical and lexical expressions, intensifiers, superlatives, or special punctuation (Mohammad, “Imagisaurus: An Interactive Visualizer of Valence and Emotion in the Roget’s Thesaurus”). The sentiment analysis approach used in the present paper focuses on the detection of so-called emotion or mood words, i.e. words that are understood as possessing an emotional and/or sentiment value. This approach makes use of a definite list of words – ideally compiled by means of a theory-driven and controlled procedure – associated with one or more values, which typically reflect whether a word embeds sentiment and/or emotions, and how intensely. Within literary studies, the theoretical frameworks most commonly used for the investigation of sentiment and emotions are respectively Russell’s two-dimensional sentiment model (“A circumplex model of affect”) – with valence as the degree of positivity and negativity expressed, and arousal representing the strength of the sentiment expressed – and Ekman et al.’s atlas of emotions (Ekman et al.) – featuring six basic emotions: anger, disgust, fear, joy, sadness, surprise. These word lists can be then mapped onto a corpus, providing a measure of sentiment and emotions. This approach is generally referred to as lexicon-based, and it is one of three broad classes of computational tools used for sentiment analysis across disciplines, which can be distinguished according to whether they use word lists (e.g., Jockers), vector-space models (e.g., Turney and Littman), or a combination of both (for reviews see Taboada et al.; Jacobs and Kinder). In line with other scholars (Jacobs et al.), we rely here on the lexical coverage of lexicons – i.e. the percentage of words of a given text or corpus covered by a given sentiment lexicon – to establish the validity of our results.4

Some limitations have to be kept in mind when working with lexicon-based approaches: the finite amount of words in the lexicons (on which the coverage often depends); the differences between one lexicon and another; and the sometime lack of transparency in the way the lexicons were developed. Moreover, sentiment lexicons are built on contemporary language usage, and might not be comprehensive of historical forms or archaic meanings of words that are pertinent to a specific historical corpus such as the one we are examining here. The two alternative approaches – creating (and validating) a custom sentiment lexicon, or using machine learning – are however extremely resource and time intensive. Since tests in the current historical project showed a word coverage comparable to contemporary language, we decided to utilise out-of-the box resources.5 Lexicon-based approaches remain among the most easily accessible and practicable methods for sentiment analysis (Jockers and Thalken; Flüh), and as these represent the essential building ground for more complex unsupervised sentiment models (Nakov et al.; Barbieri et al.), it is crucial to keep using and improving them.

The application and further development of sentiment analysis techniques has been impacted by industrial interest in customer research, with most of the research done on the English language. Not as much attention has been given to the investigation of literary and other creative texts, and up to now only few resources are available for languages other than English. Nonetheless, in recent years, with the expansion of the digital humanities, new research has resulted in valuable lexicons also for other languages, including German.

We chose to examine emotions and sentiments in our corpus using all the open source lexicons for German available to this date, as suggested by recent research in the field (Kern et al.; Schmidt, Dangel, et al.). We hope therewith to provide a ground for gauging not only how these lexicons compare to each other, but, most crucially, how valid they are for analysing the representation of emotions in relation to space in historical German-Swiss fiction (~1840-1940).

It is not the scope of the present paper to provide a comprehensive comparison of the existing sentiment lexicons for German, nor to evaluate in depth their coverage of the historical dimension of affective language. However, some of their characteristics – in particular regarding the number of words included in the lexicons and their coverage of our corpus – do help us evaluating the potential differences in their performance. A minimal summary of the lexicons used and their composition is shown in Table 2.

From the prominent role of a Swiss-national framework that marked literary production during our time frame, and that putatively sees an opposition between rural/natural and urban spaces, we formulate here two main hypotheses: (1) natural and rural landscapes will be overall more present in our corpus in comparison to towns and cities, and (2), natural and rural landscapes will be encoded overall as more ‘positive.’ Aware that realist writing does depict problematic aspects of life in rural surroundings, however, we retain that the difference in emotional encoding between urban and rural spaces may not be as neat as we hypothesise.

Our exploration will start with (1) a description of the distribution of different types of spatial entities detected in our corpus, and move on to (2) investigating their affective encoding. Importantly, this extends to questions about how ‘spatially self-referential’ the texts in our corpus are, comparing the proportion of mentions of Swiss geographical locations with those of countries surrounding Switzerland (the current territories of Italy, France, Austria and Germany).

We are not aware of any research investigating the presence and distribution of non-named spatial entities along with geolocations in fictional texts, nor systematically comparing rural and urban environments (real or fictional) in terms of the sentiments and emotions related to these. It is this gap that the present paper addresses. Using on the one hand lists of spatial entities that can be matched to our historical literary corpus, and on the other sentiment lexicons that capture the sentiment and emotional value of words, we examine here how the rural and natural landscape and geographical locations are represented, in opposition to urban space and settlements.

The compilation of a corpus of Swiss fictional narratives written between the years 1843 and 1940 presented several challenges, in particular due to the lack of a comprehensive catalogue of Swiss fictional works for those years, and to the limited amount of works already existing in a digital form. With the collaboration of experts from the University Library Basel, a comprehensive list of the authors and works of interest to our project was produced,6 featuring more than 7,000 titles of fictional narratives written by Swiss authors in German between the years 1880 and 1930.7

Using the list of authors and titles obtained, we aimed to gather as many of the listed works as possible in their digital version. A corpus of 125 Swiss texts was thus compiled, combining 79 texts which already existed in digital form and matched our list of Swiss fiction texts – from Gutenberg.de (Projekt Gutenberg DE) – with 46 additional texts, which were newly digitised by the University Library Basel as part of the research project ‘High Mountains Low Arousal? Distant Reading Topographies of Sentiment in German Swiss Novels in the early 20th Century’ as a contribution to the European Literary Text Collection (ELTeC) (Schöch et al.). The corpus contains narrative texts by 42 different authors, with a maximum of n=14 works attributed to a single author (Heinrich Federer) and a minimum of n=1 (several authors). More details about the final corpus are available on our online repository.8

All of the texts were stored as plain text files, and were then processed with the German language model of the R package ‘Udpipe’ (R Core Team; Wijffels), to tokenise and lemmatise the texts while preserving information about paragraphing and sentence structure.

In terms of the operationalisation of literary landscape, our categorisation of spatial entities overlaps in part with what Wartmann et al. (1580) identify as biophysical landscape elements – i.e. ‘terms relating to geology, landforms, soil, land cover, flora, fauna, and climate,’ as distinct from cultural landscape elements – ‘terms referring to land use, settlements, infrastructure, domesticated animals, and anthropogenic objects.’9 Central to this definition is the notion of landscape as being something which is perceived and defined by the ways in which it is described by people (Scott, “Assessing Public Perception of Landscape: The LANDMAP Experience”; Scott, “Assessing Public Perception of Landscape: From Practice to Policy”).

We found however that this taxonomy only takes into account the distinction between ‘natural’ and ‘human-made,’ but not that between ‘rural’ and ‘urban.’ Because both ‘rural’ and ‘urban’ are related to human practices, we decided to use a broader umbrella notion of RURAL for all terms related to ‘nature’ and proper names of natural places, as well as terms referring to rural landscapes and proper names of rural settlements. Analogously, URBAN is our umbrella term for both proper names of larger settlements – cities and towns – and for terms describing – or characteristic of – the urban landscape as generally conceived in the European culture.10

In order to further differentiate RURAL and URBAN space we identified four ‘entity’ subcategories for the broad category RURAL, namely ‘rural’ and ‘natural’ terms, ‘rural geolocations,’ and ‘natural geolocations;’ as well as two entities subcategories for URBAN: ‘urban geolocations’ and ‘urban’ terms. We collected as many terms as we could for each subcategory, including, where possible, terms that are characteristic of Swiss landscapes/architecture, such as Hütte or Dörfli. The six categories are described in Table 3 below:

Because of the morphology of the German language with its abundance of compound words, the identification of verbs, adverbs and adjectives potentially belonging to the spatial categories revealed complex and often ambiguous ( e.g., directional adverbs such as talabwärts, bergabwärts: ‘down the valley,’ ‘downhill,’ or berghoch: ‘as tall as a mountain / ’uphill;’ all containing ‘Tal’ – valley – and ‘Berg’ – mountain). It was therefore decided to focus on substantives only.

Rural, urban and natural terms were initially collected from openthesaurus.de (Naber), Wikidata (Wiki), and Swiss Idiotikon (Hunger et al.). These lists were then enriched with the natural terms gathered in the Text+Berg project Bubenhofer et al. as presented by Derungs and Purves and with the urban terms collected by Bologna, translated into German. The lists were then manually checked and improved with the help of a student assistant (L1 German and Swiss German), who removed redundancies and cross-category terms. While we discarded texts fully written in German-Swiss dialect from the corpus because of problems with language processing – especially in tokenization and sentiment analysis – some relevant dialectal words (e.g. Hütte, Dörfli) and as many Swiss-specific terms as possible were added for each list.

For geoloc_urb, geoloc_rur and geoloc_nat locations of today’s territories of Switzerland, Germany, Austria, Italy and France, open data available from geonames.org (Wick and Boutreux) and Wikidata (Wiki) were used. We decided to include geographical locations of countries surrounding Switzerland as these are likely to play a part in the representation of the RURAL and URBAN landscape in German-Swiss literature in the historical time frame considered. These sources allowed us to store also more specific information, in particular about the category geoloc_nat, enabling us to distinguish whether these were streams or lakes, valleys, mountains or forests.

After a process of selection, removal of duplicates, hyphenated location names, cross-category terms (for example Kirche – ‘church’) and names of places that correspond to common nouns (for example Zug, which also denotes ‘train’) or proper names,11 we obtained a final list of N=173,913 spatial entities belonging to the five subcategories of RURAL and URBAN. Details of these are summarised in Table 4.

After removing German stopwords, symbols and punctuation, we matched our lists of spatial entities onto the corpus. We created therewith a new corpus composed by the spans defined by the matched spatial terms, where each span comprised 101 words: one matched entity, the 50 words preceding it and the 50 words following it. The resulting corpus counted N=6,776,378 tokens. We then merged each sentiment lexicon onto our corpus by lower case lemmas – having also lemmatised each token – and calculated aggregated values for all the sentiment/emotions for each of text span defined by the presence of a spatial entity. Sentiment and emotion values were aggregated by span. For lexicons measuring discrete emotions, we gave a score of 1 to each word in the span with an emotion value, and counted the sum per emotion, per span, thus obtaining counts for each discrete emotion. Also for lexicons measuring sentiment (arousal, valence) we aggregated the values by span: where values were continuously scaled, we calculated the mean values for arousal and valence. Since mixed models for lexicons were computed independently, we did not need to normalise values across lexicons. If the valence value was categorical (negative vs. positive), we converted these respectively into -1 and 1 and calculated the mean value by span.

As each span was defined by the presence of a single entity, we were able to label each span as either RURAL or URBAN. This allowed us to examine the corpus as a whole, analysing statistically the two broad spatial categories RURAL and URBAN for differences between encoded emotions and sentiment values.

To establish beforehand whether the entities in our lists possessed themselves an emotional value (as this could affect the values for sentiment and emotions of the spans) we found that spatial terms covered together only the 1.1% of the entities. This shows on the one hand that the sentiment lexicons used contain very few locations or spatial terms, and on the other hand that the entities themselves – necessarily present in each text span in our corpus – would be unlikely to bias the sentiment and emotion value of the surrounding text.

For our statistical analyses, we used linear mixed-effect models (Kuznetsova et al.; Bates et al.; Baayen), with the broader space type (URBAN/RURAL) as fixed factor, and the author and title of each of the spans as random factors. We ran mixed models for each lexicon independently.

For spatial entities we report descriptive statistics in terms of raw and relative frequencies.

Figure 1 compares the distribution of spatial entities (rural, urban, natural, geoloc_rur, geoloc_urb, geoloc_nat) in the list (on the left side) with that of the corpus (on the right side). Colors distinguish between the main categories RURAL and URBAN. We observe a pronounced skewness for our entities list, where combined RURAL entities amount to 96.7% in total, a number that largely surpasses that of URBAN ones (3.3% in total).

Here, it is important to note the categorical difference between the two bar charts: In the left bar chart, we see the entity list, which was compiled from multiple resources. It is a proxy for the distribution of factual geographical locations in Switzerland, France, Germany, Italy and Austria: Here, the total number of villages, mountains, rivers, lakes and valleys – i.e. geoloc_rur and geoloc_nat entities – greatly surpasses the number of large towns and cities (geoloc_urb). In this bar chart, the number of named-entity terms is also by far higher than the number of rural, urban and natural non-named entity terms terms.

By contrast, the bar chart on the right shows the text corpus. Searching for the list’s terms in the N=125 Swiss literary texts (N=3,539,090 tokens), we find a total of N=67,093 (meaning that 1.9% of all tokens in our corpus are either RURAL or URBAN spatial entities).

Here, the difference between the macro categories URBAN (green) and RURAL (blue) is much less pronounced, but still remarkable: around three quarters of all detected spatial entities fall into the RURAL umbrella category (73%, vs. 27% URBAN).

Details of the distribution of entities in the Swiss corpus can be observed in Table 5 (for ease of comparison, the distribution of the entities list (left column) is repeated from Table 4).

In the corpus, the overall prevalence of RURAL entities (n=48,950) in comparison to URBAN ones (n=18,143) is particularly due to natural terms, the biggest category overall, with 36.5% of all spatial entities detected (n=24,500 tokens). It is followed by the urban category, with n=11,889 tokens, and the geoloc_rur (n=11,864 tokens) category, which each contribute about half the proportion of natural.

Relative to the number of terms in the entities lists, we note the highest coverage for non-named entities (natural, urban and rural). By coverage we understand the percentage of unique terms (corpus-linguistic types) per category actually found as tokens in the corpus).

This can be explained by the relatively small number of terms included in these three categories, and the fact that we are dealing with common terms, as opposed to proper names of locations. However, it is notable that the relatively small categories natural, urban and rural (n=1060 in total in the entities list) represent about two thirds of all the spatial entities found in the corpus (n=44,920, 67%).

This is quantitative support for the observation that general and common terms (our rural, urban and natural categories) are used across narratives to build the so called ‘storyworld’ (Gavins; Troscianko): creating settings, landscapes, natural phenomena and objects in the fictional world (Nünning). Geolocations (geoloc_urb, geoloc_rur and geoloc_nat), on the other hand, while making up an overpowering proportion in the spatial entities list (see Figure 1), are comparably less frequent in the corpus (33%). Such low frequency can be explained by their semantic specificity, and by the fact that fictional texts do not necessarily mention (m)any existing geographical locations. Given the sheer number of locations listed in our entities list, it is not surprising that only a small fraction would make it into literary texts, and fewer still into a relatively small corpus such as the one analysed here.

Nonetheless, when considering the percentages of geolocations (geoloc_rur 17.7%, geoloc_urb 9.3%, geoloc_nat 6%), the dominance of RURAL is again notable: geoloc_rur and geoloc_nat together amount to 23.7% of all spatial entities. Together, their substantial proportion (one fourth of all spatial entities) may be interpreted as reflecting a type of realistic writing that is anchored not only in (imaginary) references to ‘meadow,’ ‘river,’ ‘hill’, but which can be mapped onto factual geography.

As mentioned above, for each sentiment and emotion value present in the sentiment lexicons, we fitted a mixed model which looked at the difference between spans defined as RURAL or URBAN based on the spatial entities detected in the corpus. The model had thus RURAL/URBAN as fixed factor, and included title and author as random effects (formula: ‘value ~ type_grouped + (1|author/title)’).15 All the results of these models are summarised in Table 8 and 9 below. The columns RURAL and URBAN summarise the result of each model, with a ‘+’ and a ‘-’ simplifying the models outcome. When a ‘+’ or a ‘-’ is shown, it indicates that the entity type RURAL/URBAN had significant effect on the sentiment/emotion value. A ‘+’ shows an increase for the value considered, while a ‘-’ indicates a significant decrease. If the cell contains ‘na’, the entity type had no significant effect on that value.

In line with our expectations, we detected a general increase in the emotional ‘richness’ of spans containing RURAL entities, in comparison to URBAN ones. Particularly informative are the results obtained with the SentiArt lexicon, where this was true for all discrete emotions (positive as well as negative ones). With a lexical coverage of 81.8%, results obtained with the SentiArt lexicon can be considered as comparatively most reliable.

With the caveat of a low coverage, the Stamm and Klinger et al. lexicons show significant differences for discrete emotions between RURAL and URBAN as well: Stamms’ anger, sadness, and fear appear to be significantly more present in RURAL spans, as well as Klinger et al.’s anger, sadness, fear and contempt. Only Stamm’s anticipation and disgust appeared to be higher in URBAN spans, while models for all other emotions did not show any effect of space categories. Because of the low coverage of these lexicons, these results must be however interpreted with great caution.

When looking at sentiment values, we found again results in line with our hypothesis. Both BAWL-R and LANG lexicons featured a significantly higher valence in spans containing RURAL spatial entities, and, more importantly, the same was true for SentiArt’s AAP value (affective-aesthetic potential), which strongly correlates with valence (Jacobs and Kinder, “Computing the Affective-Aesthetic Potential of Literary Texts”). Along with the other discrete emotions detected by SentiArt, the AAP appeared to be strongly affected by the entity type, with spans containing RURAL entities featuring significantly higher AAP than those containing URBAN ones, thus indicating a more ‘positive’ encoding for the first.16

In terms of arousal, no difference was found for either the BAWL-R or the LANG lexicons.