Analyzing networks in communication: a mixed methods study of a political debate

Abstract

How can we detect and analyze network ties in non-reactive communication data? Previous network research mostly relies on the mere occurrence of events between actors (e.g., e-mails, citations). The paper proposes combining qualitative interpretation of communication and formal quantitative analysis. Conversation analysis and interactional sociolinguistics give us methods to qualitatively discern how actors are related in communication (relational meaning). This can be typified into kinds of relational events like “attack” or “support”. Types of social relationships (“alliance” or “conflict”) are characterized by specific combinations of relational events. These methods are used exemplarily to analyze a televised political debate between six party representatives in Germany. Three prominent types of relational events are identified qualitatively: interruptions that are supportive or adversarial, and accounts of action. Hierarchical cluster analysis shows these relational events to fall into four types of relations (alliance, all-out attack, moderate attack, ignore).

Introduction

Social network research increasingly combines formal quantitative methods with qualitative interpretation (Crossley 2010; Fuhse and Mützel 2011; Domínguez and Hollstein 2014; Herz et al. 2014; Bellotti 2015; Bernhard 2018). Most of the discussion focuses on qualitative interviews. These are seen as a way of extracting more complete information on networks from individuals, and of complementing the formal network perspective with an understanding of the subjective meaning tied to relationships and networks. However, qualitative interviews are not always possible or warranted, e.g., when reconstructing social networks from archival or web sources. Therefore, this essay pursues a different strategy, of combining formal network analysis with the qualitative interpretation of communication. The research question is: How can we reconstruct social relationships and networks from non-reactively observed processes of communication among a set of actors? This paper develops requisite methods and illustrates them by way of the analysis of a televised political debate.

The approach builds on a long-standing tradition of determining network ties on the basis of communicative events sent from one actor to another (like letters or e-mails; Kossinets and Watts 2009; Kitts et al. 2017). Much of this research follows the “formal” approach of network analysis: Events are taken to signal ties between actors without considering their content or meaning. But do e-mails, for instance, really make for similar relationships, irrespective of whether they are queries, requests, invitations, or complaints? These kinds of communication are the realm of qualitative approaches to communication like conversation analysis and interactional sociolinguistics (Schegloff 2007; Schiffrin 1994; Tannen 2004). Their methods should be picked up on to complement the formal analysis of events in networks. In this vein, McLean (1998), McFarland (2001), and Gibson (2005) examine communication in networks with regard to its meaning and multiple facets.

To reconstruct network constellations in communication the following procedure in three steps is proposed:

• (1) Documents/utterances are interpreted qualitatively with regard to how they relate the actors involved (relational meaning).

• (2) This qualitative analysis leads to the construction of types of relational events with similar relational meaning, e.g., “likes” on social media or adversarial interruptions in conversation.

• (3) Types of relational events are coded and counted in their occurrence over time and over ties. This allows discerning patterns of events characterizing the ties between actors, as well as micro-interactional dynamics of relational events.

This paper develops these methods out of the discussion of relevant approaches, and it illustrates them with a brief, rather limited empirical study. The empirical case is a TV debate of the six representatives of the most important political parties in a regional election in Germany, which is analyzed with regard to the interaction between them. The aim is to reconstruct the relations of alliance, conflict, and ignoring between the six political parties, as presented to the audience in the debate.

The paper starts by discussing the relevant methodological approaches: the analysis of communicative events in network research (2) and the qualitative approaches of conversation analysis and interactional sociolinguistics (3). Section 4 gives a methodological summary of the lessons from this discussion. The empirical example of the political debate is introduced in Sect. 5, and Sect. 6 gives an overview of the methods used. An exemplary interpretation of a short debate segment follows (7). This leads to three kinds of relational events identified in the debate (8). The quantitative analysis then groups ties with similar distributions of relational events and constructs a simplified constellation of the six parties based on their prevalent ties (9). The discussion (10) and the conclusion (11) summarize the methodological approach and highlight its potential as well as its limitations.

Networks of relational events

Network research has long relied on the network questionnaire as the prime method of gathering information on ties (Adams et al. 2020). Supposedly, social relationships are based on individual choice, and derived from unilateral nominations. Kitts (2014: 275ff) emphasizes the observation of behavior as a second method: Rather than rely on individual reports of dyadic ties, researchers can examine what happens between actors. Letters, e-mails, phone calls, citations, or co-involvement in social gatherings indicate the existence of ties (Hinde 1976). This method suggests itself for online or archival material in computational social science (de Nooy 2015; Kitts and Quintane 2020: 83f). Citations, retweets, and many other communicative events can be studied with automated methods that make burdensome network surveys obsolete. Also, the data is process-generated and non-reactive, and less prone to distortion by interview effects or missing data from non-response.

Events in networks are typically analyzed in two ways:

• (1) They are traced in their dynamics over time to discern the impact of past events and of their relational constellations on whether a particular event takes place (Butts 2008; Kossinets and Watts 2009; de Nooy 2011; DuBois et al. 2013; Kitts et al. 2017). The methods of relational event history models and of time-lagged regression models allow identifying causal antecedents of events in a multivariate, multi-mechanism framework, including homophily, reciprocity, and transitivity. Most of these studies examine only one type of event: speaker-transition, book review, e-mail, hospital-transfer of patients. The strategy is also limited in that the focus on the dynamics of interaction does not cover network ties. These would have to structure the occurrence of events over and above interactional dynamics.

• (2) Alternatively, events are aggregated over time periods to observe network ties as relatively durable regularities of communication. If two actors repeatedly interact with each other in a particular way, this indicates a particular social relationship between them (de Nooy 2015; Fuhse 2022: 29f, 254f). E.g., Papachristos measures gang rivalries through the number of homicides between gangs in three different time periods (2009). Similarly, studies of scientific citations or of Twitter retweets frequently aggregate events over time, irrespective of their temporal unfolding.

Again, most of these studies only examine one kind of event, and they do not consider the meaning of events for the relations at play. Retweets, citations, e-mails and other events relate actors to each other, but may do so in varying ways. Common-sense understandings of social relationships like “love”, “friendship”, or “patronage” suggest they combine different kinds of recurrent interaction (Fuhse 2013; 2022: 53ff). Particular interaction like empathic conversation or the exchange of material goods may occur in different kinds of ties. One type of relationship differs from another by the variety of interaction in it: Romantic love and friendship share many of the same activities, but differ in the inclusion or exclusion of bodily intimacy. Similarly, family and business ties both feature economic transactions. These are combined with contractual obligations (business) or with mutual care and the sharing of meals, household chores, and leisure (family).

In this sense, Martin defines relationships “in terms of the particular actions appropriate to it—which I shall call an ‘action profile’” (2009: 11). The definition of a social tie as friendship, love, patronage, or conflict prescribes the kinds of interaction compatible with this definition, and thereby makes for their recognizable difference. To observe social relationships empirically therefore requires, first, identifying multiple kinds of interaction. Secondly, a particular type of social relationship exists to the extent that typical combinations of these multiple kinds of interaction between two actors can be detected. These two steps allow classifying ties as “friendship”, “mentorship”, “conflict”, or “love”.

Before types of relationships, the relevant kinds of interaction have to be identified. Activities like “bodily intimacy” and “economic transactions” are easily detected. But how to discern a particular conversation as empathic, citations or retweets as supportive or critical, and e-mails as formal or personal? There is no guarantee that any two citations, retweets, or e-mails carry similar relational implications. Most of the research on relational events pursues the prevalent “structural” perspective of network analysis, and considers two events (e.g., citations or e-mails) as occurring or not-occurring at a particular point of time—as on or off, like ties in a network matrix. But what is the “relational meaning” of a particular communicative event? This calls for the qualitative interpretation of communication and suggests a combination of qualitative and quantitative methods.

The relational underpinnings of communication are not the same as its content (Watzlawick et al. 1967: 32ff; Fuhse 2022: 242–244). It is therefore necessary to distinguish the study of relational events from that of the content of communication, which has become a prominent tradition over the last 30 years (Danowski 1993; Evans and Aceves 2016). Here, a body of text is analyzed with regard to the systematic links (co-occurrences or grammatical connections) between semantic units, usually words. This leads to the reconstruction of cultural or “semantic networks” (Carley and Kaufer 1993), where relations run between symbols rather than between actors as in “social networks”. The two approaches studying social and cultural networks have recently been linked in socio-semantic network analysis (Roth and Cointet 2010). For instance, we can investigate the extent to which the hostility in parliamentary interaction between political parties in parliament reflects their ideological similarity or difference in word usage in speeches (Fuhse et al. 2020).

In linguistic terms, relational underpinnings involve the pragmatics of communication, as opposed to syntax and semantics. But of course, these aspects are not wholly separate: The question from a colleague to a conference presentation has to be examined in its content to determine whether it was supportive or critical, and how it relates the academics to each other. The task is to isolate the relational aspects of communication by way of qualitative interpretation, and to consider how the variety of communicative events with similar relational meaning is distributed across ties between actors.

Few studies of communication in networks have a closer qualitative look at these events:

• De Nooy and Kleinnijenhuis examine the dynamics of attack and support in an election campaign (2013). This requires the interpretation of the referrals between actors in statements.

• Gibson studies communicative events in manager meetings with regard to the people addressed in the first and the second turn, building on the insights from conversation analysis on turn-taking (2005). Do turns in these meetings address the previous speaker, the group, or somebody else? Different types of ties (superiority, friendship) are characterized by typical distributions of communicative events (2005: 1579ff).

• McLean examines the usage of vocabulary like “service”, “friendship”, or “honor” in letters of patricians asking for favors in Renaissance Florence (1998). He shows with qualitative interpretation and quantitative analyses that these cues correspond to the characteristics of relationships like strength and status inequality.

• McFarland and Diehl’s research on disruptions on classroom similarly combines qualitative and quantitative methods (McFarland 2001; Diehl and McFarland 2019). They analyze the effects of informal networks and of characteristics of the classroom situation on disruptive communication.

These studies point to the fruitfulness of analyzing formal networks and communication in conjunction. But they do not discuss how to identify and isolate the relational meaning of communicative events. The use of qualitative interpretative methods for this task is as yet underdeveloped.

Qualitative approaches to communication

Qualitative network interviews draw on a host of methodological approaches for eliciting and analyzing subjective meaning. Other qualitative approaches study meaning in communication. This section offers a cursory discussion of two prominent approaches: Conversation analysis and interactional sociolinguistics. Both share a common ancestry in Goffman’s dramaturgical approach.

Goffman studies face-to-face encounters in diverse situations, looking for regularities in how the participants establish some kind of common understanding (the “definition of the situation”) in their interaction. His studies range from detailed micro-analysis of encounters with participant observation ([1959] 1990) to the interpretation of communication transcripts (1981). Goffman views communicative events as geared at establishing the “footing” of the participants in a common “participation framework” (1981: 124ff). Any communicative sequence features, and builds on, a (changing) definition of how the actors relate to each other: a relational definition of the situation. Important components of this participation framework are the “faces” of the actors involved (Goffman [1955] 1967). “Face” here stands for the social identity of actors as negotiated in interaction. All of these key ingredients—the definition of the situation, the participation framework, footing, and actors’ faces—are realized and develop over the course of communication—rather than in subjective meaning.

Conversation analysis retains Goffman’s focus on micro-interaction, combining it with the ethnomethodological search for rules underlying behavior. The approach looks for regularities in the usage of linguistic forms, in particular with regard to the sequential ordering of communicative events (Sacks et al. 1974; Schegloff 2007). Turn-taking regimes come with different rules for the succession of utterances in informal conversation, teaching situations, or public debates. Every communicative event lays the ground for subsequent events, accumulating to the recursively linked processing of meaning in communication.

With the focus on the sequential order of communicative events, conversation analysis identifies types of communicative events. For example, the distinction between “questions”, “answers”, “offers”, “rejections”, and other types of events allows examining their systematic relations in typical “adjacency pairs” like question–answer (Schegloff and Sacks [1973] 1984: 73ff). As in Weber’s general method of typification, these are ideal types grouped together by their similar meaning. Weber calls for the interpretation of subjective meaning based on observable behavior. In contrast, conversation analysis follows de Saussure’s structural linguistics in locating the meaning of linguistic forms in their systematic relations to other linguistic forms. The meaning of a (type of) communicative event here lies in its sequential connection to other (types of) communicative events.

Some conversation analysts also study the meaning of linguistic forms and events sequences for the constellation of actors. They find that social relationships affect conversational practices such as introducing a topic (Maynard and Zimmerman 1984) and talking about other people (Raymond and Heritage 2006). Pomerantz and Mandelbaum (2005) examine the deployment of categories for relationships (like “friendship”) in conversation. Relatedly, stories can be told in order to promote one’s own status and denigrate others (Goodwin 1990). Particularly in conflicts with multiple parties, alignment structures and participation frameworks are constructed and negotiated with sophisticated conversation techniques, such as piggybacking (Goodwin and Goodwin 1990). Collaborative story-telling, in contrast, aligns speakers with each other (Lerner 1992). Generally, the construction of turns and the communication of agreement contribute to the construction of identities (“faces”) in relation to each other (Lerner 1996). This more “social” side of conversation analysis strongly overlaps with interactional sociolinguistics.

Sociolinguistics examines language with regard to the social context, for example the intermingled use of two languages in multi-cultural contexts (Gumperz 1982). Interactional sociolinguistics observes the impact of this context on the interaction between the people involved, linking Gumperz’s micro-studies of linguistic meaning to Goffman’s “participation framework” (Schiffrin 1994: 97ff; Tannen 2004). For example, discourse markers like “but” or “you know” link discourse units, and they position speaker and listener to each other (Schiffrin 1987).

Within interactional sociolinguistics, the politeness approach focuses on how communication deals with preserving “face” of the participants (by averting “face-threatening actions”; Brown and Levinson [1978] 1987). For example, requests are typically formulated in the mode of “indirectness” with a lot of hedges, in order to not impose and to leave the target some choice in the matter (without threatening the requester’s face).

But linguistic forms like indirectness do not only “politely” save other people’s faces. They also manage social relationships. This relational research focus has been pushed in recent years by a number of sociolinguists. Identities and relationships are seen here as “interactional achievements” (Arundale 2010). “Relationship thinking” is integral to our linguistic codes, with the relevant codes differing widely between cultural groups (Enfield 2009). The management of social relationships has become a central point of interest for sociolinguistics as “managing rapport” (Spencer-Oatey 2002) and “relational work” (Locher and Watts 2005).

Interactional sociolinguistics, with its many branches, has its core focus on the construction of identities and relationships. Many studies examine qualitatively how particular linguistic forms are deployed to affect the alignment between speakers (and audiences), relying on interpretation based on familiarity with cultural forms. In combination with the more structural conversation analysis, this points the way to a systematics of types of communicative events with regard to their relational meaning. This allows quantification and formal analysis. But already the qualitative-interpretive approach offers important insights into the micro-management and construction of relationships and networks.¹

Methodological summary

The paper starts from the consideration that networks of social relationships can be fruitfully studied in communication. Rather than rely on the self-reports of actors in interviews, this approach analyzes the actual process of interaction. The quantitative work surveyed in Sect. 2 points in this direction. It examines (1) the relational dynamics of communicative events in patterns of reciprocity, transitivity, homophily, or preferential attachment (relational event models), or (2) the regularities of communication in ties—how often do any two actors in a network population interact over time?

Both approaches rely on the clear-cut identification of communicative events. E-mails, letters, citations, or retweets are considered as discrete and countable units. These are placed in three-dimensional arrays of network matrixes of actor-to-actor ties (two dimensions) over time (third dimension). At any point in time, a particular event occurs from actor A to actor B (1) or not (0). This allows for a wide range of formal-quantitative analyses. However, the procedure relies on the reduction of the complexities of communication to simple events that can be counted, to ones (occurring) and zeros (absent). This comes with two rather basic problems of measurement:

• The events in question have to be clearly identifiable. This works well in cases of e-mails, citations, and retweets. However, already these pose problems of identification: Should the citation by A of a work by two authors B and C be counted as two separate events, one citation from A to B and one from A to C? Naturally occurring communication in face-to-face meetings or phone calls is even more messy, with difficult coding decisions for seemingly straightforward items like “turn-taking” or “question”.

• The events are counted and treated as similar in subsequent analyses. Two citations, two e-mails, or two retweets feature as undistinguishable units, even though they might be very different in content and implications. An e-mail between two students could be friendship chatter, an invitation to a party, asking for a date, to coordinate a presentation, or even harassing an unpopular dorm mate. Treating all of these instances as comparable assumes that their differences do not matter much for the research question at hand, or that they do not vary systematically. However, a glance at the list of possible e-mails above suggests that this is implausible: Friendship chatter, party invites, date requests, the coordination of presentations, and harassment are bound to occur in very different kinds of contexts and social relationships.

These arguments caution against a purely formal approach and admonish a more detailed study of communication. What is the meaning of a particular e-mail, citation, or retweet? What are its implications for the relationship between the actors involved? These questions are difficult to answer with purely computational methods, which focus on the content of messages (Evans and Aceves 2016). Here, the qualitative approaches from Sect. 3 come in. Conversation analysis and interactional sociolinguistics offer methods for analyzing communication in its fine-grained details and implications for the situation at hand. They build on the interpretation and typification of communicative events. This includes the grouping of events as similar based on the similarity in their conversational environments—as when questions are identified from typical follow-up responses, even if they do not show the expected grammatical structure.

Of particular interest for the study of social networks is the “relational meaning” of communicative events. How do types of events like questions, queries, interruptions, or invitations position actors in relation to each other? What are the implications of linguistic forms like discourse markers, indirectness, or story-telling? Conversation analysis and interactional sociolinguistics already offer a number of insights in this regard for network researchers to draw on.

Overall, then, the following steps are advisable for the analysis of social networks in communication:

• (1) In a first step, the boundaries of a discourse have to be identified. Which texts have to be included as interconnected body of work? And which actors are involved? The latter question may not be always be trivial, since communication is not always addressed and attributed to individuals. Similarly, collectives and corporate actors, including companies and states, can become cornerstones of discourse, if communication targets them rather than particular spokespersons. Some communication may even address or be attributed to multiple kinds of actors, as when a campaign statement is seen as coming from a politician and from her political party.

• (2) Then, the communication has to be scanned for prominent and recurrent ways of relating. These can occur in grammatical patterns like questions or orders, in forms of address (formal/informal you, first names, “my friend”, “buddy” etc.), in discourse markers like “but” and “y’know”, in interruptions and lower-voice backchannels (“hmm”, “yeah”), in story-telling, in use of pronouns (“we”, “our”, “you”, “they”), in support or criticism, and many more.

• (3) These ways of relating have to typified by the similarity of their relational meaning. Do a class of events show similar linguistic features, and do they have equivalent implications for the relations at hand? Apart from the features of the communication at hand, the conversational environment (build-up, reactions) gives indications for the similarity or dissimilarity of events. Preference should be given to easily identifiable features of speech. They do not only lend themselves better to identification and quantification, but also have a higher chance of being recognized and reacted upon by participants and observers of communication.

• (4) These events then have to be identified in the material and coded as occurring between two (or more) actors at a particular point in time in a three-dimensional actor to actor by time array (see above). Most often, events will be directed, as when actor A interrupts actor B. They may also be undirected, as with two or more authors publishing a paper together.

This three-dimensional data array then allows for multiple kinds of quantitative analyses:

• (5) Network ties can be discerned from the distribution of relational events through longer time segments. Importantly, multiple kinds of relational events should be considered, since types of relationships are characterized by the combination of typical kinds of communication or activities in them. These can be reconstructed by cluster analyses grouping ties with similar event profiles.

• (6) Relational event models study the micro-dynamics of interaction. Like the network ties, these dynamics frequently involve multiple kinds of events, as when a question is countered with an answer, or when the attack on a political actor triggers support from its allies. Ideally, the micro-dynamics of interaction, the regularities of communication in relatively stable social relations, and their change over time would be analyzed in conjunction. The methods for this are still missing, though.

This list shows that it will not be possible to analyze communication in its full multi-faceted relational implications and intricacies. Therefore, researchers have to do some cherry-picking and focus on a small number of communicative features. However, the qualitative steps (1) to (3) make for a more fine-grained analysis of communication with regard to its relational underpinnings. The requisite level of resolution should be determined with regard to the research question at hand.

The political debate

The second part of this paper presents a brief empirical example of the methodological procedure sketched above. The subject is a political TV debate in Germany, thus a rather limited case. The chief aim is to illustrate the methods suggested above, rather than a full analysis of the debate. The debate took place on May 2, 2012, with the candidates of six political parties standing for regional election in the biggest German provincial state of North Rhine-Westphalia (17.5 million inhabitants).² Each candidate tried to lure voters to their respective party by presenting themselves as competent and as pursuing the best policies. At the same time, they related to each other through questions and interruptions, through criticisms and support.

The debate was chosen because it includes a relatively large number of candidates (six) in a relatively open debate format. This allowed them to intervene in each other’s turns, with both support and criticism. With a length of 100 min, it offers a fair amount of material to examine, while clearly confined in communicative events and in the relevant actors. Finally, political debates exemplify how relations in communication matter. The relations between political parties represented and negotiated in such debates structure the political field (Clayman 2004). The analysis probes the methods above to reconstruct these political relations, as projected in the debate communication.

The debate took place in an important, still on-going phase in German politics: The traditional system consisted of four established parties—the moderately leftist Social Democrats (SPD), the conservative Christian Democrats (CDU; with their Bavarian sister party Christian Social Union, CSU), the “Free Democrats” (FDP) focusing on economic liberties, and the ecological, left-liberal party Alliance’90/The Greens. This system had to incorporate the post-socialist party “Die Linke” (The Left) from 1990 onwards. The Left party is a pariah in the German party system, with socialist demands and a pronounced regional identity (tied to the former GDR). In the early 2010s, the Pirate Party entered the scene. Originally formed by internet activists, they rode one a general frustration with established party politics, in particular among the young and well-educated. Table 1 gives an overview of the political parties in the debate with their ideological leanings, their representatives, their roles in the election, and the subsequent election results.

Table 1 Political parties and their representatives in the debate

From 2010 to 2012, the state government was formed by the Social Democrats (SPD) and the Greens. CDU and FDP had formed the previous state government, and they worked together on the federal level in the second government under chancellor Angela Merkel (CDU). The Leftists played the role of an established pariah party nobody wanted to form a coalition with, especially not in the West. The role of the Pirate Party was not clear, yet, as the established parties questioned its seriousness and legitimacy in the political arena. The debate was held with two moderators (Sabine Scholt and Jörg Schönenborn). The analysis focuses on the 30 directed ties between the six party representatives.

Methods

In line with the approach sketched above, the analysis focuses on the direct “relating” of the party candidates in the debate communication. This contrasts with the quantitative analysis of issue positions in debates (Doerfel and Marsh 2003), which builds on the “semantic networks” approach (Sect. 2). However, the two approaches may well be combined. The analysis follows the steps listed in Sect. 4 and adapts them to the subject matter of the political debate:

• (1) The body of communication is defined by the debate, which clearly makes for connectedness of communication and the mutual orientation of actors. The actors in question are determined as the political parties by the research question. It was ascertained qualitatively that the candidates are treated as representing these parties in the debate, rather than featuring as individuals. Communication and previous political demands and measures were attributed to their political parties.

• (2) The debate was scanned for ways of communicative relating by viewing the whole broadcast and periods of intense deliberation a number of times, until the most important kinds of events were identified. Particularly contentious debate segments with numerous interruptions were transcribed to facilitate interpretation.

• (3) The qualitative analysis revealed three types of relational events as particularly meaningful in the debate: adversarial and supportive interruptions, and accounts of actions by the other political parties.

• (4) These were coded into a data frame of directed ties and assigned speaker turns. A particular speaker turn T by party A could feature accounts of actions by the other parties B, C, D, E, and F, as well as be subject to adversarial or supportive interruptions from these other parties. The coding was done by the author and checked against the independent coding of a student assistant.

• (5) Finally, types of directed relationships between the parties were reconstructed from the data frame through Ward hierarchical cluster analysis. This revealed four types of ties with similar distributions of the three types of relational events.

Steps 2 to 5 and their results are described in more detail in the following sections.

Communicative relating in the debate

The analysis of the debate starts with step (2): the qualitative interpretation of communicative relating. As expected, the 100 minutes of the debate offer a wide range of relational positioning between the actors, little of which lends itself to formal quantification and analysis. This section offers a brief exemplary analysis of a short fragment from the debate where the controversial “Betreuungsgeld” (“attendance subsidies") was discussed (transcript 1). In 2012, the federal government got pushed by its Bavarian coalition partner CSU to introduce subsidies for parents keeping their children at home, rather than having them attended in day-care centers. In the fragment, the conservative candidate Norbert Röttgen defends the measure against criticisms from the other parties.

Transcript 1: Attendance money (01:05:26–01:05:40)³:

The 14 seconds of the transcript cover a tumultuous period, as occurring a number of times in the debate. Overall, contentious turn-taking remains exceptional and occurs in spikes. Most of the time, just one candidate speaks without interruptions, after being asked a question by the moderator (giving her or him exclusive rights to speak). Without going into detail, the following moves can be discerned:

• (1) In lines 1–2 and 8–9, CDU candidate Norbert Röttgen (NR) defends the attendance subsidies by caricaturing the position of the Social Democrats and the Greens in the state government. He argues that the two parties exaggerate the role of the day care-center in education. This makes for their insistence on all children going into day-care. These lines can be read as an account of the ideological positions and the policies pursued by the Social Democrats and the Green Party.

• (2) Sylvia Löhrmann (SL), the representative of the Greens, interrupts Röttgen in lines 3–4. She starts with “‘course” (“‘türlich”), short for “of course” (“Natürlich”), as Röttgen says “everybody” (indicated by the bracket in the transcript). Structurally, her utterance looks as if supporting Röttgen’s position (“of course”). Since Röttgen was caricaturing her position (and that of the Social Democrats), the interruption is adversarial, contradicting Röttgen as the current speaker.

• (3) In line 5, Prime Minister Hannelore Kraft (HK) of the Social Democrats tries to chip in. Given the political constellation, she probably criticizes Röttgen and supports Löhrmann. But her interjection remains unintelligible.

• (4) Röttgen interrupts Löhrmann in line 6 with “no but”. Both of these words signal opposition, even if we interpret this short interjection on its own. This validates the interpretation of Löhrmann’s interruption as adversarial.

• (5) In line 7, Sabine Scholt (SS), one of the moderators tries to take over with a firm “OKAY”. This cuts off Löhrmann's intrusion in line 4. But she lets Röttgen continue speaking as the assigned speaker.

• (6) Christian Lindner (CL) from the Free Democrats interrupts Röttgen in line 10–11. He starts by addressing him by first name, “Norbert”. In a debate where most participants (and the moderators) called each other Mr. or Ms. plus their surname, the first name constitutes an important tie-sign, signaling familiarity (Fine et al. 1984).

• (7) This tie-sign is followed by: “now I soon cannot hold out anymore.. now I soon can’t bear it anymore with the attendance money.” We can classify this as an adversarial interruption, since Lindner interrupts and signals disagreement with Röttgen as the current speaker. But the form is peculiar: Lindner first displays familiarity and then announces that he “now soon” cannot bear it anymore. Presumably, he does not want to be seen as intruding on Röttgen’s turn. He also signals an interest in voicing his opinion on the matter at greater length—something that severely violates the rights of the assigned speaker, in contrast to Löhrmann’s more modest intervention. This leads the moderator Sabine Scholt to deny Lindner’s intervention in lines 12–13.

Since Röttgen presents and criticizes the positions of Löhrmann and Kraft (in 1–2 and 8–9), their interruptions may be legitimate—as an opportunity to “set the record straight”. Additional analyses show that accounts of action frequently trigger adversarial interruptions in this sense (see Appendix 1). Apparently, these kinds of interruptions do not violate the conversational rules in the debate.

The same does not hold for Lindner, yet he wants to voice his disagreement. He can only do so with a construction that does not really express disagreement so much as pre-announce it. Conversation analysis calls this kind of move a “pre-expansion” that prepares the ground for the following utterance, similar to pre-telling like “you never guess who I met today” or to pre-invites like “What are you doing this afternoon?” (Schegloff 2007: 28ff). Lindner’s pre-announcement defers the attack to soften the blow, just as the pre-invite defers the real invite to impose less on the conversational freedom of the next speaker. The tie-sign “Norbert” can be interpreted as serving the same purpose.

This short debate sequence features a number of ways in which the speakers relate to each other. All of them show their unique properties, given the exact words used and their precise conversational environment. Nevertheless, these events are sufficiently similar to classify them into distinct types. For example, there are three adversarial interruptions from Löhrmann to Röttgen (line 3), back from Röttgen to Löhrmann (line 6), and from Lindner to Röttgen (line 10). Röttgen’s turn in line 8 cannot be counted since it does not display a clear orientation towards a particular speaker. Rather, this picks up on his own utterance in lines 1–2 (and on his interjection in line 6). The “no but” in line 6, in contrast, cannot be read as following up on his own speech. Rather, it is oriented to Löhrmann’s turn in line 3–4. This in a sense elevates her to the status of current speaker, though it was really his turn to speak.

Three types of relational events

Step 3 consists of the inductive identification of “relational events”—recurrent types of communicative events with similar relational implications. For example, the three adversarial interruptions identified above different markedly in terms of length, content, and words used. Some words like “but” or “not” indicate opposition and contradiction. Such clues are not used in all kinds of adversarial interruptions, as evident in the transcript. In particular, Löhrmann’s interjection in lines 3–4 shows all indications of supporting Röttgen as the current speaker. But since Röttgen was caricaturing her position (and that of Kraft), this apparent support becomes criticism. This example shows the pitfalls of automated text analysis, and the importance of qualitatively interpreting speech or documents up close.

The relevant types of relational events have to be determined pragmatically by ascertaining the identifiability of events with similar relational implications and the interpretation of their meaning. In the qualitative inspection of the debate, three types of recurrent and prominent relational events were identified: interruptions that are adversarial (1), or supportive (2), and accounts of action (3).

Interruptions between the participants are a prominent feature of political debates (Tannen 2012). They consist of an (if only short) overlap of talk in conversation, a violation of the basic conversation rule “one speaker at a time” (Okamoto et al. 2002). In informal conversation, such overlaps frequently form part of the smooth transitioning between speakers, with the second speaker starting at a “projected transition point” with minimal overlap without violating conversational norms (Sacks et al. 1974: 707f). In debate communication, the turns are allocated by the moderators and usually not self-allocated (Sacks et al. 1974: 701). Interruptions clearly breach the debate turn-taking system, visible to participants and audience as violations of the rules. These features make interruptions between the debate participants easy to spot. An interruption during the debate consists of one participant intruding into another participant’s turn.

• (1) Most striking about political debates is the preponderance of adversarial interruptions. Here, one debater invades another’s turn to voice disagreement.

• (2) In contrast, interruptions are permitted if they are supportive or signal interest (Tannen 1989; Okamoto et al. 2002). These tend to be quite short, spoken at lower volume, and without attempting to take the floor from the first speaker.

In political debates, supportive or adversarial interruptions are not designed to offer reassurance to the speaker, or to argue with her. Instead, they chiefly serve to signal the position of the interrupting party, and to support or to undermine the argument voiced for the benefit of the debate audience.

Some interruptions are spoken at a low voice and/or broken off after a few seconds. These short interventions remain unclear with regard to their content and relational meaning. There can also be neutral or side remarks, directing attention in another direction (for example when making a joke). Both kinds are left out of the analysis.

• (3) “Accounts of action” are a third type of relational events. These communicative events feature accounts of somebody else’s action(s). To be included in the classification, accounts have to be offered by one of the candidates on the dispositions or behavior of other debaters (or their parties). Following Weber, “action” refers to observable behavior that results from supposed underlying subjective dispositions. This connects with the sociology of “accounts” (Scott and Lyman 1968). Accounts are relatively short statements given about one’s own behavior or that of others. Accounts of action make for the construction of actors with specific dispositions as responsible for these events (Blum and McHugh 1971).

The debate features a lot of accounts of the form: “You did A out of ideological inclination B.” Röttgen’s portrayal of the state government’s policy of promoting day-care in transcript 1 is one example. As in this case, most of the accounts of actions in the debate are adversarial. In terms of politeness research, an account of somebody else’s actions constitutes a face-threatening act (Brown and Levinson [1978] 1987). Following Raymond and Heritage, accounts of action formulated about co-present others infringe on their epistemic rights (2006). People are supposed to know best what they did and why. To give an account of their actions threatens their “face”. Some accounts of action are outright “face-denigrating”.

Communicative events were interpreted conservatively and only assigned to one type of relational events if clearly falling into it. Often, turns were interrupted multiple times by the same speaker. These are coded as only one interruption, unless they pertain to separate segments of the interrupted turn. Turns with longer stories recounting more than one action by another actor were coded as one account of action. However, turns often give accounts of action by more than one actor. E.g., Röttgen sketches a joint position of the Social Democrats and the Greens in the transcript above. These are coded separately for the two ties (from CDU to SPD, and from CDU to Greens).

Overall, 137 relational events (by three types) can be identified through the 100 min of the debate between the party candidates (Fig. 1). The main thrust of interaction takes place between the two blocks of the provincial government coalition (Social Democrats and Greens) on the one hand, and the conservative-liberal coalition of CDU and FDP on the other hand. Most attacks are to be found in these eight ties. SPD and Greens form a unified block with supportive interruptions between them. In contrast, Christian Democrats and Free Democrats more often attack than support each other. The Left Party and the Pirates remain by and large outsiders of the debate, at least in these three types of relational events. There are important exceptions, though: Kraft attacks Pirate representative Paul with five adversarial interruptions and one account of action. And Röttgen attacks Leftist Schwabedissen ten times (six adversarial interruption plus four accounts of action).

Fig. 1

Relational events during the debate

Types of ties

In step 5, the distribution of these three kinds of relational events over the 30 directed ties in the network is examined. The aim is to reconstruct types of relationships with characteristic combinations of the three kinds of relational events. For this purpose, several models of cluster analysis were conducted. A solution with four clusters derived from Ward hierarchical clustering best captures the differences in the data (see appendix 3). The four clusters and their features are listed in Table 2. They are composed of 2, 3, 10 and 15 ties and differ in the frequencies and types of relational events. The names of the four clusters derive from their prevailing features.

• The small cluster “alliance” combines the two ties between Social Democrats and Greens. They exhibit only supportive interruptions and no attacks.

• The second cluster “all-out attack” groups three ties from Social Democrats to Christian Democrats and to Free Democrats, and from Greens to Free Democrats. These ties feature the highest levels of both kinds of attacks: adversarial interruptions and accounts of action, and only one supportive interruption.

• The third cluster displays moderate attacks or mixed styles of engagement. It is relatively large with ten ties. These are characterized by a fair number of adversarial interruptions (1 to 7) and accounts of action (1 to 4), but less than in all-out attacks. Half of the ties also display supportive interruptions, mildly mixing interaction styles. Most moderate attack ties run from CDU and FDP to the ruling block of SPD and Greens, and between CDU and FDP. The SPD sends an additional two “moderate attack” ties to the outsider parties of Leftists and Pirates, and the CDU moderately attacks the Left Party.

• 15 ties are characterized by widespread “ignorance” or inactivity in the relational events considered. They show very few accounts of action and adversarial interruptions, and no supportive ones. This cluster comprises almost all ties to and from the Left Party and the Pirates, apart from the moderate attacks from Christian Democrats to Leftists and from Social Democrats to both outsider parties.

Table 2 Four clusters of directed ties with distinct combinations of relational events

The list of procedural steps stops here. Many formal quantitative methods can be used to analyze the patterns of ties resulting from step 5. This particular constellation is rather confined, with only six actors and 30 directed ties between them. This does not allow for sophisticated methods like blockmodel analysis to discern role patterns, or exponential random graph models to trace network mechanisms like homophily, reciprocity, transitivity, and preferential attachment in the network.

Nevertheless, the resulting constellation can be interpreted with regard to underlying patterns. Figure 2 presents a simplified version of where the four types of tie run, with the six political actors grouped into pairs of two by their roles in the debate: SPD and Greens support each other in a mutual alliance, sending mostly all-out attacks to CDU and FDP. These respond with only moderate attacks, which they also aim at each other. Left Party and Pirates are mostly ignored, and do not send attacking or alliance ties to others. Only the SPD moderately attacks them. Singular ties of one type, like the moderate attack from CDU to the Left Party, are dropped here in favor of the predominant ties between blocks (in this case, “ignorance” from CDU and FDP to Leftists and Pirates).

Fig. 2

Types of tie (clusters) between three blocks of actors. Gre: Greens; Left: Left Party; Pir: Pirates

These three discourse positions of SPD and Greens, CDU and FDP, and Left Party and Pirates, respectively, correspond to their institutionalized roles of state government, established opposition (and federal government), and outsider parties. These political roles seem to come with particular “discourse strategies” (Gumperz 1982), which by and large govern the relational events in the debate. Whether this pattern resurfaces in other debates remains to be seen.

Discussion

The empirical analysis of the political debate mainly serves to illustrate the combination of qualitative and quantitative methods to study network constellations in communication, as outlined in the first part of the paper. Its scope is rather limited, but it gives an overview of how to use these methods, of their potential, and of their limitations. The key results are:

• The qualitative identification and typification of three types of relational events: supportive interruptions, adversarial interruptions, and accounts of action;

• The quantitative grouping of these types of events in four types of directed relationships between the political parties by way of hierarchical cluster analysis: alliance, all-out-attack, moderate attack, and ignore.

Qualitative examination of the distribution of these four types of ties among the six parties shows their constellation to mainly follow their institutionalized roles in the political realms. The six actors by and large interacted with each other on the basis of their positions of state government, established opposition, and outsider parties.

As in the ethological studies of relations among animals (Hinde 1976), it was possible to reconstruct the web of social relationships out of the qualitative observation of interaction. Rather than ask actors for their ties to others, this approach considers what is happening between them. The transformation of the messy process of communication to network constellations rests on three key steps: the identification of kinds of relating (1), their typification as relational events (2), and the reconstruction of kinds of relationships (3). All three are fraught with problems:

• (1) The identification of relational events cannot possibly capture all forms of relating in communication. It may not even find the most important ones. Researchers have to rely on their own interpretation of the text, preferably working in teams to improve the analysis.

• (2) The forms of relating then have to be grouped into types of relational events based on their similarity and dissimilarity. Here, a number of less prominent and rarely occurring events have to be dropped, even if they are quite consequential for the discourse at hand. E.g., a political debate might feature an insult from one party representative to another. This insult would affect and mark their relationship more than all interruptions. However, its one-of-a-kind nature would not make it amenable to quantitative analysis. Also, the analysis could err by ignoring important kinds of relating, by needlessly splitting relatively similar events, or by grouping events together based on superficial similarities. Here, the typification can only work on the basis of insights from the qualitative approaches to discourse from Sect. 3, from past studies on the kind of discourse (here: political debates), and from careful examination of the material at hand.

• (3) The reconstruction of types of social relationships relies on quantitative methods like cluster analysis to group similar cases. However, these methods do not tell researchers unequivocally how many clusters to identify. In the case at hand, a solution with three kinds of relationships would also have been possible, but only with the loss of considerable heterogeneity (see Appendix 3). Researchers have to find a pragmatic solution between capturing as much covariance as possible and presenting a clear-cut constellation of low complexity.

Do the relational events and the kinds of relationships identified in these steps constitute “real types” (McKinney 1969), recognized and oriented to by the participants? Or are they more or less arbitrary research constructs? It is generally wise not to overestimate one’s analysis by claiming to have found something “real”. Both the types of relational events and of relationships may reflect gradual tendencies or more fine-grained distinctions in the real world. This does not necessarily make the analysis and the types less valid, as long as they adequately map empirical observations.

Conclusion

The methods suggested allow reconstructing network constellations from communication. This gives researchers a source of network data that is quite different from network surveys. The approach is applicable to all kinds of discourses with multiple actors relating to each other, as in political debates or parliamentary interaction in multi-party systems or among multiple candidates in a primary, but also in academic or intellectual deliberation, in face-to-face meetings in organizations, and in online interaction. Of course, the chief focus lies on the constellations of actors at play, rather than the topics or ideas discussed. The qualitative interpretation and identification of relational events delimits the amount of text.

On the plus side, network analysts do not have to rely on self-reports on social relationships, which frequently dramatically diverge from each other. Also, missing data due to unavailable or unwilling interview partners is not an issue here. The approach focuses on what is going on in communication. This opens up archival and historical data, for example from parliaments or intellectual discourses, as well as process-generated data from the internet for analysis (de Nooy 2015; Fuhse et al. 2020; Kitts and Quintane 2020; Lewis 2021). For all these, it is difficult to impossible to run network surveys.

The chief downside is that this approach does not have interview respondents translate their messy relationship communication into quantitative data. Researchers have to perform this important step now. Qualitative interpretation and typification of communication is burdensome and time-consuming, and they come with their own sources of error. Human coding always has a certain variation, collapsing very different instances into broader categories. Biernacki (2014) therefore admonishes relying on the “humanist interpretation” of textual data alone. This, of course, entails giving up on formal quantitative analysis and runs counter to the ambitions of this article.

Alternatively, Lee and Martin (2015) propose countering the problems of human coding by relying on quantitative text analysis. This would avoid the unreliability of human coders to arrive at seemingly objective mathematical representations of communication. In general, the content of texts lends itself better to the study by numbers than their relational underpinnings. The latter have to be carefully investigated by recourse to cultural vocabularies for relationship messages and metacommunication. It might be possible to qualitatively examine a smaller part of a larger corpus, and then to develop computational methods for identifying the kinds of relational events in quantitative text analysis. One version of this is the usage of sentiment analysis to detect positive or negative referrals (Featherstone et al. 2020). However, quantitative analysis should be flanked by qualitative validation. The interpretation of the brief debate transcript in Sect. 7 suggests that it is also necessary to examine the conversational environment, in addition to textual cues in the communicative events themselves.

Of course, all of these problems can be avoided by simply counting events like e-mails, retweets, or citations. This is commonly done in network studies of relational events, as surveyed in Sect. 2. But it ignores the many different facets of these events—e-mails are not always friendly, retweets not always supportive, and citations not always approving. This may work well in many instances, but misses important aspects of communication and social constellations (Keuchenius et al. 2021). Social relationships are commonly understood to be characterized by combinations of activities. They should be reconstructed as typical combinations of relational events, rather than only from one type. This calls for disentangling different kinds of relational events in communication, and for putting them back together in characteristic packages of relational events. This paper argues for the use of qualitative methods to discern relational meanings of communication. They can well be combined with automated quantitative text analysis, or the two methods can mutually inform each other (Mohr et al. 2015).

One particularly promising road ahead lies in the dual examination of discourse with regard to its content and underlying social constellations as socio-semantic networks (Roth and Cointet 2010). This approach studies the distribution of symbols, ideas, or other cultural items in a population—the “semantic network”—in tandem with the network of social relationships (see Sect. 2). The methods advanced in this paper offer a more nuanced and fine-grained reconstruction of different kinds of relationships from discourse alone, rather than relying on network surveys. This can easily be combined with quantitative text analysis like topic models and scaling techniques. This way, it is possible to combine multiple dimensions of discourse in the analysis: content and social relations, as well as their temporal dynamics.

Appendix

Appendix 1: Interrupting accounts of action

Many of the interruptions in the debate occur when or after another participant’s party’s actions were discussed. This appendix tests the hypothesis that formulating an account of others’ actions makes them likely to interrupt. On the one hand, the actors probably see the need to react to accusations and false portrayals. On the other hand, accounting for another actor’s actions violates conversation rules. It runs against the epistemics of social relations, and it often involves a direct attack on the face of the actor in question (see 8.). This recounting of another person’s actions would allow them to intervene and offer their own perspective—in spite of the strict turn-taking system of debates.

The following analysis only covers longer speaking turns of the debaters (not the moderators) of more than 10 s. 45 percent of these turns (32 out of 71) formulate an account of another debater’s actions. Often, these accounts feature more than one fellow debater (or her or his party). For example, Röttgen gives an account of the joint actions of Kraft (Social Democrats) and Löhrmann (Greens) in their coalition in transcript 1. Of course, he is interrupted by them.

To study the impact of accounts of action and other aspects of the communicative and social environments in more detail, a mixed effects logistical regression was conducted with the occurrence of adversarial interruptions from one debater to the current speaker in any particular turn as the dependent variable (Table 2). All 71 turns can be interrupted by five other debaters, which makes for 355 cases of potential interruptions. The independent variables include:

• Whether an account of action is formulated about the potential interrupter

• Whether the speaker had interrupted the debater in question in their last turn

• The length of the current turn (in seconds)

• Random effects for the 30 directed ties; this controls for the overall propensity of the actor in question to interrupt the current speaker in the debate

• Random effects for the 71 turns to control for the general level of contention in them.

The results of this mixed effects model are reported in Table 3. A debater will interrupt somebody else’s turn with a 20 times higher probability when the subject of her or his account (p < 0.001). Controlled for this, speaking for one second longer increases the odds of being interrupted by a mere 0.72 percent (not significant), and having been interrupted by the current speaker in the previous turn makes it 2.5 times as likely to interrupt their turn. This reciprocity effect is considerable, but does not reach significance (p = 0.12) when controlling for the level of contention per tie.

Table 3 Mixed effects model of adversarial interruptions occurring per turn

Both the effects of reciprocity and of accounts of action are instances of the interactional effects of previous communication on the occurrence of current relational events. The reciprocity effect only pertains to one type of relational events. In contrast, the effect of accounts of action is an example of the micro-dynamics of interaction linking different types of relational events. Its magnitude and significance provide evidence for the contention around these accounts of action. The observed pattern is a general feature of this debate: accounts of action lead to interruptions by the subjects of these accounts. To a certain extent, these interventions seem to be normatively allowed—“normal”—given that a negative account of action severely threatens somebody else’s face. Therefore, interruptions should not really be attributed to the interrupting actor alone. Rather, they serve as an indicator of what is going on in the relation between speaker and interrupter.

The random effects control for ties and turns. The effects of ties can be interpreted as signs of relationships as regularities of communication over and above the micro-dynamics of interaction. However, the random effects for the 30 ties are relatively minor and do not add much to this particular model. This probably means that the conversational rule behind the main effect operates across social relations: Attacks in accounts of action make interruptions permissible to correct one’s negative face projected by the current speaker. At the same time, social relations play out in these typical sequences of relational events. Some ties feature a lot more of them, others none. And these differences do not surface in this model which mainly captures processual regularities.

Appendix 2: Original transcript and notes on the translation

Given the problems of translation, the text offers only a rough translation of the original sequence (transcript 1). Many features of speech like intonation and emphasis are difficult to keep in translation, if there is no one-to-one correspondence of syllables and word order. Intonation and emphasis are not interpreted, and they are not included in the transcript. Here is the original German sequence:

Appendix 3: Cluster analysis

Hierarchical cluster analysis groups cases by their similarities on a set of attributes. The cases here are the 30 directed ties between parties. The attributes consist of the numbers that the three types of relational events occur in these ties. Ward cluster analysis does this on the basis of the minimum variance method, where cases are grouped together if displaying the minimal error sum of squares—minimal variance between them. The cluster analysis was performed in R with the hclust function and yielded the following dendrogram (Fig. 3):

Fig. 3

Ties are numbered with the first number for the sending party, and the second for the receiving party.

Ward hierarchical cluster analysis of ties with solution of four clusters.

The red boxes mark the four cluster solution chosen, with from left to right: “moderate attack”, “all-out attack”, “ignore”, and “alliance”.

The heights in the dendrogram mark the dissimilarity of ties or clusters based on the minimal variance criterion. The longer the lines, the more heterogeneity in the counts of relational events is lost by grouping ties or clusters. For example, the third cluster in the dendrogram assembles the “ignore” ties with very similar counts of relational events (all close to zero). In contrast, the other three cluster feature similar internal variance. In the next step, the two clusters “moderate attack” (left) and “all-out attack” (second from the left) would be grouped together, but with considerable heterogeneity between them.