(2.1) Trapped in the string query? Collecting a corpus of articles containing “Europe” from the impresso app

The recent evolution towards the digitisation of historical sources, especially the press, makes it possible to deal with the discussion of Europe in a broader sense rather than focusing on dedicated organisations or in dedicated publications. Using the digitised newspapers opens the possibility to study the broader social uses of Europe beyond the formalised ideas of Europe, developed by organised entities (Abel, 2013; Ackerley, 2019; Koenen, 2018; Wijfjes, 2017).

Contrary to many other cases, we benefited from easy access to the digitised newspapers via the impresso interface, which enables us to explore digitised newspapers from Switzerland and Luxembourg, mainly in German and French. This app was developed by an interdisciplinary team of computational linguists, digital humanists, designers, and historians, and it includes natural language processing (NLP) tools for historical print media and an interface for the active exploration and critical analysis of newspaper corpora.

We focused on six French language Swiss newspapers: Journal de Genève (JDG), Gazette de Lausanne (GDL), Express (EXP), l’Impartial (IMP), Le Confédéré (LCE) and l’Essor (LES). These titles cover a relative diversity of newspaper types: the Journal de Genève (JDG) and Gazette de Lausanne are considered liberal-conservative, and, although not formally affiliated to any political party, are perceived to be opinion newspapers, whereas Express and Impartial were originally created to disseminate commercial news and advertisements. Le Confédéré had close ties to the liberal party and l’Essor, to the Christian socialism (Clavien et al., 2015).

Using the keyword “Europe” can of course not be equated to the study of the idea of Europe (Grunewald & Bock, 1997), let alone, the antimodern discourse on Europe. However, as both dimensions (antimodernism and europeanism) are described with many polysemic words and are difficult to query as such. Also, in order to increase the discovery, we decided to collect a wide range of press articles and narrow down with the help of automatic annotation.

The outcome of the query for the words “Europe”, “European”, brought together with the query for “europ*” (including the variations “Europe”, “européen”, “européenne” etc. in French) for these 6 newspapers is 183 270 articles, for the period ranging from 1900–1944, with some significant difference in size, across the different newspapers titles (Figure 2). Looking at the simple frequency measure, one can already notice the increase over time of the results, which may not necessarily translate into a more intensive use of the word Europe in the press but may be connected to a change in the quantity of pages printed, to a changing quality in the OCR recognition and the subsequent higher number of hits. The later may especially explain the variations around 1917–1920, when the paper quality decreased in general all across Europe, and therefore also the quality of the digitised newspapers (Gooding, 2016). Beyond these formal biases, we observe that the frequency meets our expectations, with a peak in the early 1930s, which may reflect the debate that took place around Aristide Briand’s proposition (Fleury et al., 2001) or the 1938 peak, which may echo the political tensions in Europe, resulting in many mentions of the word “Europe”. These observations are a first and superficial check but cannot lead to conclusions. The impresso app lets us extract the articles’ bibliographic metadata, content, and detected named entities. We lose, however, the visual dimension, although it still available via the app. For instance, the caricature shown below will only appear as a short text in our corpus (Illustration 1).

This first step involving the collection of the textual source material leaves us with a large collection of individual articles but with little information on their content. We will now discuss how to add a new layer of automated annotation to navigate this collection of digital primary source material, and doing so, creating another kind of context for this historical source material.

(2.2) Past the polysemy: Finding “Europe” in the digitised press

Looking closer at each of the European projects reveals many gradations of richness: those of contradictions, unexpected ideological affiliations, and ambiguous formulations serving a particular political goal, which have since then been disconnected from the texts that has entered the pantheon (Schirmann, 2007, p. 108). Using a query with the word “Europe” returns many different types of uses of that word, and very few of them will be relevant to our research question. To sort out this richness, we resorted to topic modelling, which “models a comprehensive representation of the corpus by inferring latent content variables, called topics” (Maier et al., 2018, p. 94). Topic modelling produces an estimation of word co-occurrences in a collection of documents and then uses the co-occurrences to cluster the documents, and it has been often discussed in the context of exploring large digital textual collections covering a period of a few years (Galen & Nicholson, 2018; Jacobi et al., 2016; Mimno, 2012). For each cluster, a list of words describes the most frequent co-occurring words. We have used a popular tool, MALLET, which offers probabilistic topic modelling. This means that the clusters created by the topic models are not stable, and so, if run several times, the soft clusters which are created will not overlap. However, after a few trials, it appears that some trends in the topics (but not the exact article distribution) are repeated. One strength of this tool is that it is less sensitive to faulty text, which is a common feature of digitised historical materials.

As formulated by Baumer et al (2017, p. 1399), “[d]ocuments and words are then mapped to a low-dimensional latent space wherein geometric proximity matches human notions of semantic similarity”, capturing both the allure of this tool and the repeated calls for caution surrounding their use in digital humanities. We use it here to get a sense of the topic diversity and select articles that are relevant to our research. After a few times experimenting with the number of topics or the chronological segmentation of the corpus, we trained the topic modelling for each newspaper title separately and set the number of topics to be detected to 100 in order to have a fine-grained structuration for each article collection. We then selected the articles that contained at least 30% of a given topic. The computed topics link very divergent numbers of articles: some link thousands of articles and some just a few. We chose this threshold for a first exploration, to have enough articles for each topic in order to estimate the intra-topic semantic coherence but not too many, so that manual inspection remains doable. For the manual selection of articles tagged with the most relevant topics, the threshold was lowered to 20%.

We proceeded to label each topic. By doing so, one can also manually create connections between the topics computed on each title. This exercise indicated that the dominant topics, in quantity, are topics that consist mainly of stop words that were not removed in the pre-processing, followed by topics labelled as financial news, sports, and cultural programs. This indicates a significant number of articles containing mere mentions of the word “Europe”, without discourses on Europe. The topic modelling was intended to collect articles that correspond to expected themes. It can also be used to explore aspects that were not anticipated. In each title, we could identify topics that may raise interesting questions on the use of the term “Europe”. For instance, several topics seem to gather articles that contain reports on ceremonial speeches, or political, cultural or sporting events. This is a way to cluster articles to search for a particular use of the word “Europe” in public speeches, providing emphasis and references to relevant elements of collective imagination.

Some topics seem to indicate more than an association of themes, and even give some indication of the tone of the articles (Hulden, 2019) and contain some traces of anti-modern references (names or keywords), made possible with the setting of bigrams for the training of the topic modelling (shown below with the underscore linking two words). Following Jacobs & Tschötschel (2019), we use topic modelling to also feed directly our discourse analysis on Europe. We will now present some examples of topics produced, in their original French. The topics read as follows: abbreviation of the newspaper, T + the topic number, and the top 20 words or bigrams identified for the topic. What is interesting about these topics is the association of terms: war with peace associated with civilisation, nation and moral, offers a notable connotation, in contrast to an association with international institutions or front and army.

• GDL_T26_l’homme civilisation libéralisme l’individu doctrine humaine problème moderne valeurs reynold philosophie morale pensée dieu christianisme notion conception nietzsche mystique spirituelle
• IMP_T85_guerre peuple paix peuples nos monde nations liberté pays nation avons droit patrie lutte l’humanité vie hommes sommes civilisation l’histoire
• LES_T63_maître christianisme sois disciple fidèle_jusqu’à jusqu’à_mort shaw sois_fidèle non-résistance mort_disciple maître_sois opportuniste morales_spirituelles l›absolu transition période_transition l›eglise_demain notions_morales moude vos_ennemis

The labelling and the manual inspection of the clusters created by the topic modelling helped us to have a first global understanding of the large article collection but also detect some niches of potential anti-modern discourses, thanks to the chosen level of granularity. We could start to collect individual articles to feed a naïve Bayes classifier that would help us find additional similar texts.

(2.3) Translating the anti-modern figures into wordlists with a Naïve Bayes Classifier

The idea of using a naïve Bayes classifier as a way to transpose onto the digital material, the traditional historians’ practice of “finding and collecting similar texts”, came thanks to a fruitful exchange with Milan van Lange, during his fellowship at the C²DH (van Lange, 2019). Following extensive discussion on the research question, Milan van Lange proposed to use such a classifier and prepared a script in R for this. This tool is often best explained through its popular use in training a spam filter but has also been implemented in the detection of literary genres (Long & So, 2016). To train the naïve Bayes classifier, we collected about one hundred articles labelled as “anti-modern” and one hundred that were either labelled neutral or “modernist”. The manual collection was mostly validated by another historian, who was given the criteria of Antoine Compagnon, however this validation was conducted only on a small portion of articles, due to the limited availability of other domain experts.

The naïve Bayes classifier produces two main outcomes: a list of words that have been identified to discriminate one category from another, and an estimation of the probability of each subsequently submitted document to belong to one or the other category. This wordlist is a useful tool to reflect on the selection criteria we applied. It makes it possible to become aware of distinctive words, as determined by our own collection, not an indisputable anti-modern vocabulary list. For instance, as shown in the table below, the output of probabilities that some words are discriminatory for a given category, for the training on anti-modernism, as shown in the table below.

For example, the word “décadence” appears in the training set, i.e., the list of articles classified manually as anti-modern vs. non anti-modern, and occurs most frequently in the articles that have been tagged as anti-modern. The consequence is that the unknown, i.e., the not yet tagged articles that are submitted to the naïve Bayes classifier, containing the word “décadence” will have a higher probability to be tagged as “anti-modern” by the classifier.

Following the settings defined by Milan van Lange in his script, numbers were not excluded from the text in the pre-processing steps. Keeping numbers in the text can prove to be useful if a date is a marker of a particular discourse. Here for anti-modernism, the French Revolution being a target, we hope to catch some reference to it in our NBC net. Indeed, the bigram “révolution_1789” has the probability of 66% to indicate that a text containing it is of anti-modern discourse on Europe and of 52% to indicate an anti-modern discourse with a simple mention of Europe, based on the training set we prepared.

We can see also in the Table 1 that words like “marxisme” or “phénomène” are also used to tag texts as “anti-modern”, which reflects our selection but does not inform about the context in which they are used in the untagged articles. The fact that “Marxism” describes “anti-modern” texts reflects the importance of anti-communist texts in the manual selection, however, it is the whole list of words that is considered to calculate the probability of belonging to the “anti-modern” category. We opted for another cumulation of annotations, from this same tool, to create a more nuanced picture. This accumulation of signs of similarity (or their absence) is a way to compensate for the inherent fuzziness when using this tool for this purpose. The categories were paired with the category “neutral” for the article selection. In this context, “neutral” means a diversity of articles simply mentioning Europe, without any identifiable discourse on it. The categories were also paired with “opposite” and “similar” categories, to highlight the words that would surface for each category.

A first round of naïve Bayes classifier was used to test the reliability of this tool, which was validated after a manual inspection of samples of the results. A second round added another category: the presence of anti-modernism was separated from the expression of a European anti-modern discourse. The opposing categories to “anti-modern” were also expanded: to contrast with the target discourse, we collected training sets for reports on European diplomatic events, or the simple mention of Europe, without any discourses on it. Several categories were tested during this round, and some, like “studies on Europe” or “paneurope”, were either not yielding relevant results or the training set was not large enough to be used reasonably. A third round was conducted with a set larger than 100 articles, which resulted in a more diverse but less relevant results. A fourth, and so far, final, round focused on contrasting the selected categories to learn more about how they differed in terms of words and distribution. Instead of looking at opposing categories (anti-modern vs. non anti-modern), we contrasted the categories that were also close to each other (anti-modern vs. anti-communist), to see how it would impact the wordlists. The juxtaposition of the words produced by the pairing of the anti-modern conceptual category with “similar” categories such as the fascist conception of Europe (FE) or anti-communism (AC), and with “neutral” or “opposite” categories such as the crisis in Europe (CE) or utopian conception of Europe (UE) help us highlight the terms, persons, and slogans that we have indirectly selected with the collection of articles.