Topic models indicate textual aboutness and pragmatics: Valuation practices in Islamophobic discourse

1. Introduction

In this paper, we address the following research question: How can we identify textual and discursive patterns of (e)valuation in a medium of political discourse? We approach this question in a per se (e)valuative environment: the Islamophobic online blog PI-News (Dreesen and Krasselt; Krasselt and Dreesen). The German-language blog comprises statements with a strong negative attitude against liberal norms, especially against a multi-ethnical society and migrants. We assume that PI-News is a self-stabilizing discourse order based on (e)valuative postings addressing a largely homogeneous Islamophobic audience of readers and commentators. For the purpose of this paper, we concentrate on texts mentioning the Quran, in order to work with a thematically very focused corpus (assuming that (e)valuation always refers to some kind of object).

To answer our research question, we used a data-driven approach to give insights into textual evaluation practices within Islamophobic discourse. The paper provides two new insights: First, this research is based on John Dewey’s distinction of “valuation” and “evaluation” (Dewey, Theory of Valuation). We show how an obviously negative attitude towards the subject matter of the Quran is transformed into discursive structures of value judgments on a textual level. Second, we show how to innovatively use topic modeling – a data-driven linguistic approach – to identify evaluative practices in political discourse. We use topic modeling in a twofold way, extending its application in the digital humanities: Building on the distinction between aboutness and pragmatic macrostructures introduced by (van Dijk, “Text and Context”), topics are categorized with a specific view on linguistic practices, especially evaluation. Here, we focus on the evaluative practice of quoting. In addition, we explore the combination and pathways of topics in order to highlight how the evaluative practice of quoting unfolds in the course of texts. Finally, we discuss how quoting can be understood as an element of evaluation.

2. Background

2.1. Valuation and evaluative practices

Linguistic research on evaluation needs a theory of valuation. This sounds like a given, but it is not. Earlier linguistic research on evaluation refers to speech acts and conversational maxims (Sager; Zillig). Today, linguistic approaches to evaluation focus on practices and positioning in speech and text. A decidedly linguistic discussion of valuation issues can be found, for example, in the meta-examination of language ideologies, the affective relation of stance taking, and in the understanding of the linguistic market (Bourdieu 651–55; Kroskrity 497; Silverstein 223, 252–56; Woolard 740–41). The latter is an example of the rare adaptation of a value theory to linguistic issues. These exceptions include Keller and Alba-Juez and Thompson, who at least mention selected metaethical positions. It is noticeable, however, that theoretical and empirical studies of evaluative expressions and language action mostly lack a systematic reference to existing evaluation theories from, e.g., philosophy, sociology, or psychology (cf. e.g. Sandig; Linde; Baker et al.).

Since the 1980s, interest in the scientific study of evaluation and valuation has increased. This is also related to the fact that the development of certain valuation practices in Western societies has led to the emergence of “Valorisierungsgesellschaften” (‘valuation societies’), which are now being investigated (Reckwitz 14). Valuation studies and the “sociology of valuation and evaluation” are not only concerned with economic analyses and their criticism (Lamont; Doganova et al. 87–88); particularly striking are the analyses of the cultural valuation of symbolic goods and social practices (e.g., art, sports, music) and of value regimes (Lamont 203). Lamont asks:

“What can be done to ensure that a larger proportion of the members of our society can be defined as valuable? […] Addressing these questions will help us understand the impact of dominant definitions of worth and cultural citizenship, as well as their implications for xenophobia, racism, solidarity toward the poor, and attitudes toward welfare redistribution.” (202)

Valuing and devaluing creates and maintains social orders, which is why they need to be examined in both local practices and macro-structures in order to understand their forms, functions and effects.

The understanding of valuation in this paper differs from the aesthetic or economic understandings of taste judgments, such as those found in linguistic analyses of evaluative use of language. It is not about the use of language to implement judgments of taste and similar practices, but about an instrumental understanding of verbalized practical judgments aimed at guiding actions and problem solving. In this sense, Dewey argues for a pragmatic valuation theory. In pragmatism, action as a practical orientation is generally given priority over reflection, because human behavior is primarily shaped by issues of usability. The focus of pragmatism is how we can carry out our everyday investigations in a self-directed and productive way. Considered in this way, the reasons for (e)valuative Islamophobic practices lie in problem solving, not in, for example, ethical, aesthetic, or political speculative reflection.

This metaethical reflection of Dewey is based on his differentiation of valuation as attitude and evaluation as judgement:

“Dewey’s term ‘valuation’ covers both valuing and evaluation. Valuing, prizing, and esteeming denote ‘affective-motor attitudes[,]’[,] with more emphasis on ‘motor’ than ‘affective’. Valuing is a matter of loving or hating, liking or disliking something, where these attitudes involve tendencies to act”. (Anderson)

This means that one does not need to have an idea of what one (does not) value(s) in order to value (e.g., food, city, religion), and that valuing can be implemented through habits (e.g., buying a type of cheese) (cf. Anderson).

“Value, whether immediate or contributory, may be found without judgement, without implying cognition. If immediate, we prize, cherish, esteem, directly appreciate, etc., and these words denote affectional or affecto-motor attitudes, not intellectual ones. So we use objects as means, treat them as useful, without judgement.” (Dewey, “The Logic of Judgments of Practice” 4)

Dewey claims that we practically value without reflecting and judging, for example, by finding objects useful or not (e.g., using an umbrella in the rain).^[1] In contrast, value judgments arise in contexts in which we ask ourselves whether we should value something. In contrast to most philosophical positions, value judgments are practical judgments for Dewey: The function of a value judgment is to continue an action when the normal course of action has been interrupted by a problematic situation (e.g., considering buying an umbrella while standing in the rain). They are made by us to orient ourselves in situations and contexts (Anderson). Value judgments are “about the value of actions and objects as means—that is, their value in relation to their consequences, or the consequences of valuing them in the situation at hand” (Anderson). In addition, value judgments are tools for valuing whether and how changes to the evaluation process and objects are needed.

According to Dewey, we can recognize evaluative processes with means and ends in the use of propositions. This is useful for understanding Islamophobic texts:

“When the contexts are taken into account, what emerges are propositions assigning a relatively negative value to existing conditions; a comparatively positive value to a prospective set of conditions; and intermediate propositions (which may or may not contain a valuation-expression) intended to evoke activities that will bring about a transformation from one state of affairs to another. There are thus involved (i) aversion to an existing situation and attraction toward a prospective possible situation and (ii) a specifiable and testable relation between the latter as an end and certain activities as means for accomplishing it.” (Dewey, Theory of Valuation 13, emphasis original)

Dewey’s pragmatism assumes that head, heart, and hand are integrated in behavoir, “in which, to use more technical language, prizing and appraising unite in direction of action.” (Dewey, Theory of Valuation 65). By distinguishing between valuation and judgment, Dewey shows their mutual reinforcement. Valuation is transformative, as he explains, e.g., in The Logic of Judgments of Practice (4-9):

“appraisals affect our immediate valuings of things. […] Practical reasoning does not merely generate new appraisals; it transforms our prizings. This is the point of Dewey’s theory of criticism and taste. Judgments of the merits of prizings feed back onto our primitive prizings and transform them. They not only make these prizings more articulate (a union of prizing and appraising); in making us more vividly aware of the features of the object that we prize, they alter the directions of our prizings”. (Anderson)

The fact that the appraisals change as a result of the realization of devaluing judgements shows the relevance of the debate on Islamophobic statements in the public discourse. The statements as judgments indicate the deliberately emphasized characteristics of the evaluated object, the Quran. This offers insight into what the authors of the Islamophobic blog actually consider to be the problem with Islam and the Quran. These two aspects, the transformation and the named features, are discussed in this paper.

What is most convincing about Dewey’s pragmatism for dealing with evaluative language use is, first, its focus on practical judgements that aim to guide action and problem solving; second, its focus on the transformation between verbal estimation and evaluation in relation to a concrete object: the Quran. Let us assume that in the Islamophobic worldview Islam represents a central problem in the idealized community of, for example, an extremely homogeneous, purely Christian or libertarian-authoritarian ‘problem-free’ world. We do not know the everyday situations of Islamophobic persons. We do not know when and how Islam is habitually devalued in an Islamophobic way or when this is disturbed, and value judgments become needed. What we can look at, however, is how the Quran is evaluated in the Islamophobic discourse order of the blog PI-News. The patterned and thus habitual writing of text of this kind on the blog involves valuation as attitude and judgement with considerations of consequences. How this can be captured linguistically is presented below.

2.2. A text linguistic approach to topic modeling

How can linguistic practices of evaluation be operationalized from a corpus linguistic perspective? In the context of this study, we were interested in a corpus-driven, quantitative approach to evaluation. More specifically, we operationalized evaluative practices on the textual level by means of word co-occurrence patterns. To that end, we approached the task by applying a machine learning algorithm, namely topic modeling. Contrary to alternative corpus linguistic methods (e.g., sentiment analysis and other lexicon-based approaches), topic modeling has the advantage of working without pre-defined hypotheses regarding the linguistic form of evaluative practices. Furthermore, it is a “within-corpus” method, i.e., it does not require a reference corpus that is used for comparison or as some sort of baseline (as would be the case with keyword analysis). In order to use topic modeling for the identification of evaluative practices, we refer to a text linguistic theory, allowing to explain the output of the topic modeling algorithm by means of semantic and pragmatic macrostructures.

In the literature on topic modeling, the definition of the term “topic” remains vague and mostly alinguistic. E.g., Blei states that “[…] topic models resemble […] the thematic structure of the collection” (79) and that topics are “themes that run through them [collections of documents, JK]” (77), without defining what a thematic structure or a theme is. Blei and Lafferty refer to the topic structure of a corpus as “the index of ideas” (71) that is present in the texts of the corpus. From a purely algorithmic perspective, topics are probability distributions over the vocabulary of a corpus: topics are lists of words which have a high cooccurrence probability in documents. Thus, topics are calculated on the linguistic surface and are based on the cooccurrence of words, which is then used to “identify the aboutness of a corpus” (Murakami et al. 244). Murakami et al. point out that topics vary in the degree to which this aboutness is visible, which becomes evident when topics are annotated with a mnemonic label during the research process. Topics containing many grammatical words are harder to label than topics containing concrete nouns, which is very well explainable since concrete nouns refer to objects. The following table exemplifies this with two topics from the corpus used in this study. The aboutness of topic 18 can be summarized straightforwardly with the label islam critical movements. In contrast, the aboutness of topic 24 is considerably harder to grasp from the top words. In the literature, topics like 24 are sometimes referred to as boilerplate/noisy topics", i.e., topics with “no substantive meaning” (Maier et al. 108; see also DiMaggio et al. 586).

In an everyday understanding, the term “topic” similarly refers to the aboutness of a text. A topic can be understood as the answer to the question: “What is this text about?” (van Dijk, “Text and Context” 131). The answer to this question can take different forms, ranging from a single word to a short phrase or multiple sentences (Stede 88). Thus, the answer to the question would very likely contain words which are amongst the top words (see section 4.1) of a topic in a topic model. Similarly, competent speakers find it easy to link a text to a discourse: “Our linguistic behavior shows that we can say that a discourse, or part of it, was ‘about’ something. That is, we are able to produce other discourses, or parts of discourses, expressing this ‘aboutness’, e.g., in summaries, titles, conclusions or pronouncements in any form.” (van Dijk, “Text and Context” 131).

Theoretically more well-founded considerations of the term topic, however, can be found within text linguistic theory. According to Hellwig, the topic of a text is to be considered as something new, the answer to a question: “Ein Thema ist etwas Fragliches, zu dem in einem Text eine Lösung mitgeteilt wird oder mitgeteilt werden soll. Jedes Thema kann in Form eines Fragesatzes verbalisiert werden.”^[3] (3). Similarly, Lötscher considers the topic of a text as a gap or something controversial. e.g., an event, person or proposition (99). More recently, Adamzik proposes a typology of topics, referring to sentence semantics and semantic roles (215). Common to all approaches is that a topic is conceptualized as an abstraction or reduction of the content of a text.

An approach how this abstraction can be worked out analytically can be found in van Dijk (“Text and Context”). According to van Dijk, a topic of a text (he uses the term “topic of discourse” and “topic of conversation”, van Dijk, “Text and Context” 119) is derived in a process of abstraction a reader or listener has to make during reading/listening. This process of abstraction can be described in terms of semantic operations, by which the topic of a text can be derived from its individual propositions by specific rules, e.g., deletion, selection, generalization, and integration (van Dijk, “Text and Context” 143). Understood in this way, a topic is the result of a hierarchical organization of the propositions of a text. With focus on action and structure (see below), individual propositions in a text build macrostructures, which in turn build further macrostructures and so on, until the stage of the global macrostructure of a text is reached.

We follow van Dijk’s theory of macrostructures and argue that topics (i.e., the result of the topic modeling algorithm) are a corpus linguistic approach to this theory. Topic modeling certainly is not suited as a cognitivist model, i.e., it tells us nothing about the way a topic can be inferred from a text cognitively (e.g., by a set of rules applied by a reader/listener), but the theory of macrostructures is nevertheless a suitable framework for enhancing the analytical power of topic modeling. Topics can be understood as indices of macrostructures – above the level of individual propositions but still below the level of the global macrostructure of a text. Topic modeling follows the assumption that a text always consists of multiple topics; for each text, the proportion of each topic is calculated by the algorithm. We argue that the global macrostructure of a text can be operationalized as the combination of topics. Pathways of topics can give insights into macrorules, i.e., the way how two (or more) topics unfold linearly in a text is informative about the way these macrostructures are connected.

Furthermore, van Dijk distinguishes between macrostructures with a semantic function and macrostructures with a pragmatic function. Semantic macrostructures consist of propositions that transmit semantic information and create reference to objects:

“We see that the notion of ABOUTNESS is not very precise, and, at least for sentences, not always decidable. […] More in general, aboutness should be established in (con-)textual terms, perhaps in such a way that a discourse or a passage of the discourse is about something if this ‘something’ is referred to by most phrases with topic function. In this case, however, we no longer deal with the topic of a sentence but with a TOPIC OF DISCOURSE”. (van Dijk, “Text and Context” 119)

We prefer van Dijk’s term aboutness (see also “‘content’, ie a propositional base” 11) instead of semantic macrostructure. Pragmatic macrostructures, on the other hand, refer to the communicative function of a text and determine the interpretation of utterances (70). Following Searle, pragmatic macrostructures consist of (conventionalized) actions that may or may not succeed (van Dijk, “Text and Context” 92).

We would like to emphasize that texts are complex and ambiguous, so we do not try to map or capture textuality with digital tools alone nor a simplistic concept of text. Our approach reflects the following aspects: First, digitality reduces the complexity of communication (Nassehi). The use of machine learning is not intended to have a mapping effect, but rather to extract a few aspects from the mass of information that are considered relevant. Second, what is convincing about the idea of macrostructures is the fuzzy concept of aboutness:

“a story may be about Romeo, about Juliet, about both, about a specific (forbidden or impossible) love or about certain political structures in the middle ages. Often, however, the ‘aboutness’ pertains to a given individual object or person, if most properties and relations are assigned to one permanent referent or to those objects/persons introduced in relation to it.” (van Dijk, “Text and Context” 119)

Therefore, it depends on the discourse or media in question what a text is about (i.e., Islamophobic blog about Quran).

Third, the linguistic use of topic modeling allows to understand topics as indicators of macrostructures, not topics as propositions or practices. Topics must be interpreted in the context of the texts in which they appear. This is why we distinguish between topics indicating aboutness (“aboutness topics”) and topics indicating pragmatics (“pragmatic topics”).

Summarizing, we argue that these two types of macrostructures can also be identified in topic models, i.e., that topics indicate semantic as well as pragmatic characteristics of texts in a given corpus. Though van Dijk introduces semantic and pragmatic macrostructures as a dichotomously identifiable set of textual characteristics, we argue that in an empirical, data driven approach such a clear-cut distinction is not appropriate. Instead, topics incorporate both types of macrostructures, but to a varying degree. I.e., topics can be arranged on a semantic-pragmatic-continuum. Whether a topic is placed more towards the semantic or pragmatic end depends on the characteristics of its top words (e.g., topics with a high proportion of concrete nouns are more semantic than pragmatic), its unfolding in the text and, even more important, on the epistemological perspective under which topics are examined. In the study presented here, we analyze topics with a view on evaluation (cf. “key discursive themes” in Islamophobic blogs, Ekman 1992).

3. Data

The blog PI-News was founded in 2004 and is one of the best-known German-language websites in the field of right-wing populism and anti-Islamism (Schiffer; Dreesen and Krasselt; Krasselt and Dreesen). In a broader sense, Islamophobic attitudes can be characterized as “group-based enmity”, such as homophobia, anti-Semitism, and racism (Heitmeyer et al.; for a detailed discussion on Islamophobia in contexts of Western myths cf. Kumar 41–60). PI-News has been monitored by the German Federal Office for the Protection of the Constitution since April 2021.^[4]

The website gained popularity in connection with the controversy over the Muhammad cartoons in Jyllands-Posten in 2005. The blog received more attention in 2015, when Germany took in Syrian refugees and the Islamophobic movement Pegida (Patriotische Europäer gegen die Islamisierung des Abendlandes, ‘Patriotic Europeans against the Islamization of the Occident’) was organized. The ideological concept of assumed “Islamization” is key to almost all Islamophobic European movements: The blog describes itself as “Pro-American” and “Pro-Israeli” and advocates “Against the Islamization of Europe and For Basic Law and Human Rights”.

The blog posts used in this study were collected in a web crawling procedure in October 2021. The corpus covers a period of 15 years, with the earliest texts published in 2006 and the latest texts published in October 2021. Only texts in which word forms with the morpheme koran ‘quran’ occur (e.g., Koran ‘quran’, Koranschule ‘Koranic school’) were considered in the analysis (cf. section 1). The corpus contains N = 4,840 texts (5 Mio. words).

4. Method

Topic Modeling is a collective term for a series of probabilistic machine learning methods that identify patterns of word usage in large collections of text (Blei 77). Originally developed in natural language processing, topic modeling has become a method widely used in Digital Humanities, e.g., for the analysis of literary genres (Schöch), themes and motifs in poetry (Navarro-Colorado) and novels (Jockers and Mimno).

For the analysis of evaluative practices, we applied a three-step procedure. First, we calculated an LDA topic model and categorized the resulting topics in aboutness and pragmatic topics based on van Dijk’s theory of textual macrostructures (cf. section 2). Secondly, we examined frequent combinations of topics (with a focus on evaluation) to examine how aboutness and pragmatic topics are intertwined. Thirdly, we analyzed the pathways of topics to see how macrostructures typically emerge in texts. The subsequent steps will be explained in more detail in the following subsections.

4.1. Step 1: Identification of Topics

In topic modeling, a topic is a probability distribution over the vocabulary of a corpus (Blei and Lafferty 73). Topics are usually represented by those twenty to thirty words that have the highest probability of occurrence in that topic. Typically, these top words are used for describing, categorizing, and interpreting the identified topics (cf. section 2.2., table 1). Crucially, topics are directly linked to the documents in the corpus: each document is a mixture of multiple topics and exhibits an individual proportion of each topic. E.g., we can filter the corpus for documents in which topic 1 is most strongly represented. Each word (except stopwords, see below) in a document is assigned to a topic.

For modeling the PI-News corpus, Latent Dirichlet Allocation (LDA) was used (Blei and Lafferty). It is one of the most frequently applied topic modeling algorithms. In LDA, the documents of the corpus are treated as the observed variables which are used to infer the hidden topic structure that generated the texts (Blei 79). It is an unsupervised machine learning method, i.e., no training data is used to infer the topics. Instead, the inference of topics is performed via Gibbs Sampling, which is a specific form of Markov chain Monte Carlo (Steyvers and Griffiths), i.e., each word in each document is assigned to a topic in an iterative process.

In LDA, several parameters have to be set by the researcher. This concerns the number of topics, Dirichlet priors and the usage of a pre-defined stop word list. For the analysis presented here, the number of topics was determined qualitatively by evaluating multiple topic models with a varying number of topics. Criteria for evaluation were semantic coherence of top words and topic distinctiveness (Evans 2). The final model contains 30 topics. We further set an asymmetrical concentration parameter alpha, which allows for highly frequent words to end up in a small number of topics instead of being spread over all topics (making them semantically less coherent) (Wallach et al. 6). Additionally, a stop word list was compiled, containing words which occur in more than 30% of the documents. We did not work with a pre-defined stop word list, since such lists may contain words crucial for identifying pragmatic macrostructures and practices of evaluation in particular (e.g., quotation marks, modal particles, personal pronouns or numbers). Also, words occurring in less than 10 documents were deleted in a preprocessing step, since hardly any distributional information is available for these.

All topics were manually categorized into aboutness and pragmatic topics (cf. section 2). For this purpose, ten documents in which a particular topic was most strongly represented were examined qualitatively to see how the top words of the topic are textually embedded. In an open coding procedure (Breuer), topics were grouped based on WH-questions (e.g., in relation to events, institutions, places, objects, i.e., what, when and where) and with a view on possible practices of evaluation.

4.2. Step 2 and 3: Topic combinations and pathways in texts

After obtaining a topic model and categorizing the topics for the PI-News corpus, we identified the two most dominant topics for each document and their pathways in the text of this documents. To analyze how individual topics are used throughout the course of a text (topic pathways), the topic-word-assignment in each document was used. Each text was split in ten equally sized bins. For each bin the mean proportion and standard deviation for the two dominant topics in the text was calculated (cf. Table 3). Topic pathways are visualized by simple line graphs (one line for each topic), where the course of a text is represented by ten bins on the x-axis and mean topic proportion on the y-axis.

5. Results

5.1. Topics (Step 1)

Topics were categorized based on the idea of macrostructures indicating aboutness and pragmatics, introduced in section 2. Topics of the aboutness type were further categorized according to their aboutness, i.e., thematically similar topics were grouped under the same mnemonic label. The following topic categories were identified in the coding process (a comprehensive list of all topics is provided as supplementary material).

5.1.1. Aboutness topics

Most of the topics (24 out of 30 Topics) were categorized as aboutness topics. In these topics, mainly nouns, adjectives and verbs with a very specific meaning represent the top 30 words. E.g., topic 11 contains the nouns Schule ‘school’, Kind ‘child’, Schüler ‘pupil’, Lehrer ‘teacher’ and the adjectives muslimischen ‘muslim’, christlich ‘christian’, staatlich ‘state’. It was labeled with the mnemonic label ‘alleged Islam culture and how it contradicts with western culture’.

The categorization of aboutness topics into coherent groups remains fuzzy to a certain amount. E.g., topic 26 could also be categorized as “violence and terror with alleged Islamic origin”, since texts with a high proportion of this topic are on domestic violence. But since alleged domestic violence in Muslim families in these texts is viewed from the perspective of western values, we have chosen the category “alleged Islam culture and how it contradicts with western culture”. Despite the limitation due to their fuzzy nature, these topics can be clearly separated from the following pragmatic topics.

5.1.2. Pragmatic topics

Six topics were categorized as pragmatic macrostructures. The top words of these topics differ from the other topics, since they mostly have either a very general meaning (e.g., heißen, Teil, gelten, verstehen), are grammatical words (e.g., ein, welche, daß, dafür) or indicate the reproduction of direct quotations (topic 9: euer, vers, sure, numbers). Most of these words occur with a high token frequency which explains why they end up in the same topic(s). This is a result of our decision to set an asymmetrical Dirichlet prior Θ, which allows for individual topics to occur with a higher probability then other topics (Wallach et al. 8). Indeed, three out of these four topics (6, 24 and 25) are the topics with the highest overall probability (cf. Fig. 1). Nevertheless, these topics should not be excluded from analysis since the usage of words with a rather general meaning can reveal insights about characteristics of texts.

The qualitative examination of topics listed in Tab. 5 shows pragmatic macrostructures that indicate a hermeneutic process (topic 6, 16, 24, 25). The hermeneutic process is an essential part of value judgment in contrast to valuing. The value judgment becomes conscious when someone asking himself/herself whether something should be (de)valued in order to solve the ‘problem Quran’. This is what Dewey called the process of investigation, which includes devaluing, ignoring, and emphasizing aspects (Dewey, Theory of Valuation 50). Therefore, making sense through interpretation is at core an evaluative practice that can be identified by evidence, arguments or simply semantics on different linguistic levels. The present investigation is proceeded by the following: The topics consist of communication verbs (heißen ‘to mean’, erklären ‘to explain’, fragen ‘to ask’, erzählen ‘to tell’, bezeichnen ‘to name’, lesen ‘to read’), verbs used in the illocutionary speech act of assertives (verstehen ‘to understand’, gelten ‘to apply’, kennen ‘to know’, glauben ‘to belief’, versuchen ‘to try’) and the consideration using part-whole relations (teil ‘part’, gegenüber ‘opposite’, seite ‘side’, jedoch ‘however’, während ‘whereas’).

The hermeneutic process includes also semantic objects with and about which interpretation takes place: Communication nouns (thema ‘topic’, aussage ‘statement’, beispiel ‘example’, bild ‘image’) and other nouns that refer to different aspects (muslime ‘muslim’, deutsche ‘german’, politiker ‘politician’, gewalt ‘violence’, gefahr ‘danger’, problem ‘problem’, kind ‘child’, Freund ‘friend’, angst ‘fear’, haus ‘house’, geld ‘money’, straße ‘street’). Typical for the PI-News-Blog are truth-related nouns (ideologie ‘ideology’, faktum ‘fact’, wahrheit ‘truth’) due to the self-understanding as a counter-medium against the ‘suppressing system of lying press’ (Lügenpresse). The hermeneutic process works by focusing on some aspects while not using others in the same way for sense-making. This is supported by using words for reinforcement (einig ‘agreed’, natürlich ‘of course’, wirklich ‘really’, genau ‘exactly’, wichtig ‘important’, sogar ‘even’, richtig ‘right’, nie ‘never’, klar ‘clearly’, völlig ‘totally’, totalitär ‘totalitarian’).

Topics 9 and 15 refer to a pragmatic macrostructure which is characterized by the practice of quoting. More specifically, topic 9 is the result of extensive quotation of Sura verses (top words are Sure ‘sura’, Vers ‘verse’, the pronouns euer ‘yours’ and du ‘you’ refer to the addressing of potential readers, the numbers 4, 5 and 2 represent specific suras/verses). Topic 15 is the result of longer quotes given in English (e.g., a speech given by Barack Obama).

Quoting suras is a practice to present the subject matter to be evaluated, the Quran, in a supposedly unbiased way. Reference to the source of the quotations (www.islam.de, Central Council of Muslims in Germany) and the proof of the—translated—original sentences lend meaning to the quotation practice. Crucially, a certain image of the Quran is demonstrated – and a certain image of Islam and the collective group of Muslims is directly derived from it. For the Islamophobic writers of the blog, the Quran presented in this way is part of the problematic situation in which the evaluation process takes place. In the following section, the way Sura quotes are used in the blog posts in general will be introduced in more detail.

Figure 1.Frequency of topics in study corpus (number of texts where a specific topic is amongst the top 3 topics).

5.2. Topic 9: Quoting of Suras

Example (1) shows how quotes of Suras are typically embedded within a text. The two quoted Sura verses (2:191, 2:193) are referred to as “timelessly valid orders” given to Muslims to use physical violence against non-Muslims. The current violence exerted by members of the Islamic state is directly connected to the Quran.

(1) Diese Moslems im Islamischen Staat wiederholen exakt die grausamen Taten ihres „Propheten" Mohammed. Es wird mit den seit 1400 Jahren laufenden Massenmorden niemals aufhören, solange Mohammed als perfektes Vorbild gilt und diese zeitlos gültigen Befehle im Koran stehen: 2:191 Und tötet sie, wo immer ihr auf sie stoßt, und vertreibt sie, von wo sie euch vertrieben haben; denn die Verführung zum Unglauben ist schlimmer als Töten. 2:193 Und kämpft gegen sie, bis es keine Verfolgung mehr gibt und die Religion allein Allahs ist. (PINES_55007)

"These Muslims in the Islamic State are exactly repeating the cruel deeds of their ‘prophet’ Muhammad. It will never stop with the mass murders that have been going on for 1400 years, as long as Muhammad is seen as the perfect role model, and these timeless commands are written in the Koran: 2:191 And kill them wherever you come across them, and drive them out from wherever they have driven you out; for seduction to unbelief is worse than killing. 2:193 And fight against them until there is no more persecution, and the religion is Allah’s alone." (translated with DeepL, 2024/01/22)

The analysis of frequently occurring 3-grams in Sura quotes supports the observation that PI-News most dominantly quotes Sura verses which refer to physical violence. In total, the corpus contains 744 quotes of Sura verses.^[5] The following patterns occur in these 3-grams (table 6, highlighted in gray): (1) usage of verbs and nouns referring to physical violence (schlagt ‘beat’, getötet ‘killed’, gekreuzigt ‘crucified’, gemetzel ‘slaughter’), (2) usage of imperative constructions with plural pronouns (schlagt ‘beat’, tötet ‘kill’, meidet ‘avoid’, lauert ‘lurk’), (3) reference to non-Muslims through the noun Ungläubige (‘disbelievers’).

The high frequency of individual 3-grams indicates that Sura verses are quoted multiple times. Table 7 gives an overview over the ten most frequently quoted Suras in the study corpus. The analysis shows that only a small number of different Suras are quoted (the Quran in total contains 114 Suras), with sura 8 and 9 being the most often quoted ones (more than 100 quotations).

Table 8 gives more detail about the five most frequently quoted verses within these suras alongside an example quotation from the corpus. Without exception, all five verses shown in the table contain incitements to physical violence. In the corpus, we find multiple variants of one and the same verse, i.e., quotations are not identical in wording. For the majority of the quoted suras, the source of the German translation is not given. In a small number of cases, the website of the Central Council of Muslims in Germany is named as a source.

5.3. Topic Pathways (Step 2)

We identified topic 9, quotations of Quran suras, as a topic with a specific evaluative function. In step 2 of the analysis, we used the topic modeling output to examine how sura quotations are combined with other topics to elaborate further on this practice of evaluation. Topic 9 most frequently occurs in combination with the topics listed in table 9 (n > 20 texts).

Figure 2.Pathway of Topic 9 (blue line in each facet) in texts with topic 19 (A), topic 13 (B), topic 28 (C) and topic 26 (D), red line in each facet.

Overall, the practice of quoting Sura verses shows a preference for specific thematic contexts. A striking pattern is that Sura quotations are used in the context of violence and terror (topic 13, 19 and 28). Topic 26, which has been categorized as “islam in western society and culture” is also closely related to violence, namely domestic violence in Muslim families. Overall, this complements the findings of the 3-gram analysis for all the documents regarding the close relationship between the topic of violence and quoted suras verses.

Fig. 2 shows the pathway of topic 9 in texts with topic 19 (A), topic 13 (B), topic 28 (C) and topic 26 (D). In each facet, the blue line represents topic 9 (Sura quotations), the red line one of the other four topics. The x-axis shows the position in the text (10 bins in total), the y-axis shows the mean topic proportion (MTP). All four facets show a striking similarity regarding the distributional pattern of topic 9: at the beginning of the texts, the MTP is lowest, i.e., the first bin (= the first 10% of a text) is the one with the fewest words belonging to topic 9. In the subsequent course of the text, the MTP for topic 9 is increasing before finally decreasing again in the last bin. In contrast, the respective other topic in focus is most dominant in the first bin (with an exception for topic 26 on domestic violence, cf. facet D). Overall, the topic pathways indicate that Sura quotations are not given immediately at the beginning of a text, but later on, when the topic of violence and terror has already been introduced.

The following corpus excerpt gives an example how topic pathways instantiate in a specific text. It shows text PINES_42440 in which topic 9 (bold face) and topic 19 (on violence and terror with alleged Islamic origin, underlined) are the two most dominant topics:

Fig. 3 shows the pathway of topic 9 in texts with topic 8 (A), topic 6 (B), topic 24 (C) and topic 25 (D). In general, topic 9 shows the same pattern as in Fig. 2: its mean topic proportion is lowest at the beginning of a text, rises towards the middle of the text and decreases at the end of a text. I.e., we observe a stable pattern regarding the preferred position for Sura quotations in texts. Topics 8, 24 and 25 show a different behavior then the topics in Fig. 2: topic 8 on the role of the Quran (Fig. 3, A) has a pathway parallel to topic 9 which can be explained by their close relationship: both are closely connected to the Quran and develop in sync. The curves for topic 24 (Fig. 3, C) and 25 (Fig. 3, D) run relatively flat (i.e., without salient peaks) which is an additional argument for their categorization as pragmatic topics. They refer to a specific form of hermeneutics (cf. section 5.2.1) that runs throughout the text.

Figure 3.Pathway of Topic 9 (blue line in each facet) in texts with topic 8 (A), topic 6 (B), topic 24 (C) and topic 25 (D), red line in each facet.

Discussion

In this paper, we have shown that topic modeling is a suitable tool for the detection of macrostructures in texts. Through the decidedly linguistic application of topic modeling, aboutness macrostructures and pragmatic macrostructures can be distinguished quite clearly from each other. The aboutness topics relate to the situation in Germany and the West (topic no. 1, 2, 3, 4, 5, 7, 8, 11, 12, 17, 21, 26, 27, 30) and to (international) violence and terror (10, 13, 14, 19, 22, 28, 29). However, there are also clearly self-referential anti-Islamic (18, 20) or meta-discursive media-related topics (23). The pragmatic macrostructure consists first of topics indicating a hermeneutic process (6, 16, 24, 25). Hermeneutics, interpretation, “investigation of things” is what Dewey calls inquiry, of which value judgments are a part (Dewey, Theory of Valuation 50). The blog writers refer to the Quran and point out through different aspects and contexts the Quran as a problem in a given situation in Germany and Europe.

The second type of pragmatic macrostructure concerns the quotation topics (9, 15). It has been shown that the quotation of suras (topic 9) can be recognized as a distinctive topic. The quotations are used to specify the ‘problematic situation of Islam in Germany and Europe’ and therefore the reason for the value judgements. Following Clark and Gerrig, focussing on specific aspects and illustrating them are the very function of quoting: “Demonstrations and descriptions are fundamentally different methods of communication. Demonstrations depict their referents—what is being demonstrated—whereas descriptions do not.” (764) If we understand value judgement as a process of investigation, the quotation can be seen not only as a negative aspect of the quoted Quran as a whole: “What the speaker ultimately calls attention to is not the displayed token itself in all its singularity but certain properties of it, i.e. some type which it instantiates.” (Recanati 640). That is the reason why the quoted Suras have in some cases the textual functions ‘opening’ und ‘closing’, shown by the combination of the quotation topic (9) and the semantic topics (19, 13, 26, 28). Recanati explains:

“More often than not, the properties in question are demonstrated because they are properties of something which one attempts to depict through the demonstration. Let us call that thing the ‘target’. […] The speaker therefore does three things at the same time: he displays a token, demonstrates certain properties of that token (a type), and thereby depicts the target.” (Recanati 642 emphasis original)

The target could be ‘Islam as violence’, demonstrated by the Sura quotes in the context of violence and terror (13, 19, 28, and also 26). This finding is also supported by the analysis of 3-grams frequently occurring in Sura quotations. Following Dewey, the target is one of the most important features of the prized object ‘Quran’.

What is the status of value judgements on Quran in the Islamophobic PI-NEWS blog? The aboutness topics, the hermeneutic process, and the quotations have the function of confirming what Dewey calls the ideal:

“The ideal is, in the first place, a method of insight into the given or present situation; and primarily it is not something to be attained. That is, it is not a goal in an external, fixed sense which is there and which we simple recognize and aim for. But it is rather a way of interpreting the given experience. […] You cannot eliminate the fact that in the projection the ideal does go beyond present experience. You cannot get the ideal out by a merely mechanical study of existing conditions. But it must grow out of negative elements in the existing situation taken in relation to the positive elements.” (Dewey, Lectures on Ethics, 1900 - 1901 60)

We do not know the ideal of the Islamophobic writers in detail, although we can think of it in broad outlines: ‘Germany and Europe are no places for Islam.’ The appraisal of an object is “an act that involves comparison” (Dewey, Theory of Valuation 5). Islamophobic writers imagine a state in which the Quran plays no role. The depicted target, e.g., in quotation contradicts this ideal. What we can now see, based on the topic modeling with regard to the selected Sura quotations, is how a verification of the ideal is made: With the help of the (for the most part unspoken) ideal in the texts, the writers embedded quotations in aboutness macrostructures with violence topics to show negative effects of Islam. In this respect, the ideal is a method to find and demonstrate negative evidence of the Islam underlying the situation under investigation. Finally, the textual macrostructures transform the Islamophobic evaluation judgement of the Quran into a primitive negative prizing, called hate, again.

What is dangerous about this is that it is almost impossible to distinguish between means and ends in the valuation process: The also often unspoken means form the content of the concrete end, not some abstract norm or ideal (Dewey, Theory of Valuation 48). Moreover, ends become means for further ends (Dewey, Theory of Valuation 50). This raises the question of whether devaluating Islam and Quran is an end in itself, or a mean to achieve another end. This also concerns the legal issue how an end is to be achieved (Dreesen and Krasselt). With the costs of achieving the end, the assessment of the value of the end comes into focus: “Practical judgment is creative: it institutes new ends-in-view.” (Anderson).

Dataverse repository: https://doi.org/10.7910/DVN/FRYPSL