CSS Seminar

Textual corpora are usually not static over time: as new documents get published (e.g., news, scientific articles, tweets), topics of interest may change. Describing their rise and fall over time has generated substantial research over the last decade. Better, it turns out that considering the temporal dimension of textual modeling improves the automated description of these corpora.
Over the years, researchers of the ERIC lab have developed several models to explore this paradigm. Early works simply run static models on different time slices. More elaborate approaches consider that models estimated on each time slice are not independent from each other. Even more elaborate approaches go further, get rid of time discretization and model time as a continuous variable along with textual content.
Our talk will present an overview of these approaches, also illustrating our lab’s recent progresses in this regard – especially in terms of tracking topics over time or for studying how pieces of information interact to trigger new information on social media.

To openly express one’s views is a fundamental right in any liberal democracy. However, in the age of social media, the questions of what is allowed to say and how public discourse should be regulated are ever more contested. We present a pre-registered study to analyze citizens’ preferences for online hate speech regulation. We devise synthetic hate speech posts mimicking actual cases of hate speech. Respondents are asked to judge the posts with regards to perceived offensiveness and hatefulness as well as to actions that should be taken by the platform providers and other consequences the sender of hate speech should face. The experiment is embedded in nationally representative online panels in the US and Germany, which allows us to analyze context- as well as individual-level determinants for online hate speech regulation preferences.

In 1751, Jean le Rond d’Alembert had a dream: “to make a genealogical or encyclopedic tree which will gather the various branches of knowledge together under a single point of view and will serve to indicate their origin and their relationships to one another”. Three hundreds years later, the digitalization of human activities and its outputs has made text-strings a universal medium for addressing knowledge production in domains as diverse as academia, biomedical research or online citizen debates. We will develop the notion of “branches of knowledge” envisionned by d’Alembert and propose a reconstruction method that makes it possible to reconstruct, from unstructured textual materials, a multi-level and multi-scale representation of these branches of knowledge. We will then investigate the new relationships that this type of reconstruction allows to establish with collective phenomena by using several applications of this method such as scientific publications, Twitter data or clinical trials.

In recent decades, the effect of social networks on culture has been both confirmed and questioned. Our paper argues this puzzling contradiction to result from theoretical and methodological gaps between different accounts of networks and culture—and offers a bridge across these gaps. On the theoretical side, we propose an integral model of symbolic construction that combines the previously separately studied ego and dyadic social network effects on culture across two major social network domains: instrumental and expressive. Namely, we argue that formation of group’s vocabulary of cultural elements (cultural consensus) is driven by personal instrumental social ties, while relations between the cultural elements constitutive of their meaning (cultural tightness) owe to dyadic expressive social ties. On the methodological side, we open up an opportunity to juxtapose and disentangle the different dimensions of social network effects on culture by introducing multilevel exponential random graph socio-semantic network modeling. This method is capable of estimating the relative contribution of the two social network dimensions to culture within the same model, thus controlling for their interplay. Applying our method to a unique longitudinal multi-source dataset obtained through five two-year-long ethnographic studies of European artistic collectives, we find support for our theoretical model across creative groups embedded in diverse social and cultural contexts. The approach and our findings open up new opportunities for the sociology of culture, social network analysis, symbolic interactionism, studies of creative collectives in arts, science, business, activism, and beyond.

The rise of social media and computational social science (CSS) has opened countless opportunities to explore social science questions with new forms of data and methods. However, CSS research on socioeconomic inequality, a fundamental problem in sociology, has been constrained due to the lack of individual-level socioeconomic status (SES) measures in digital trace data. We propose a new approach to address this problem. Following the sociologist Bourdieu, we argue that the commercial and entertainment accounts that Twitter users follow reflect their economic and cultural capital and hence, we can use these followings to infer the users’ SES. Inspired by political science approaches to inferring social media users’ political ideology, we develop a method that uses correspondence analysis to project official Twitter accounts and their followers onto a linear SES scale. Using this method, we estimate the SES of 3,482,657 Twitter users who follow the Twitter accounts of 339 supermarkets and department stores, clothing and speciality retailers, chain restaurants, newspapers and news channels, sports, and TV shows in the United States. We validate our estimates with data on audience composition from the Facebook Marketing API, self-reported job titles on users’ Twitter profiles, and a small survey sample. The results show reasonable correlations between our SES estimates and the standard proxies for SES: education, occupational class, and income at the aggregate level and weaker but still significant correlations at the individual level. The proposed method opens new opportunities for innovative social research on inequality on Twitter and similar online platforms.

Multivariate networks consist of a vertex set with at least one type of edge between pairs of vertices and with numerical and/or qualitative attributes on the vertices and the edges. These networks provide a more accurate representation of social structure than univariate networks, but analysing them introduce technical and computational complexity. In this presentation, we consider exploratory and confirmatory analysis of multivariate networks represented as multigraphs. Multigraph data structure is described with examples of their natural appearance, together with a description of the possibility to obtain multigraphs using blocking, aggregation and scaling. Two random multigraph models are presented and several statistics under these models are derived. It is shown how these statistics can be used to analyse local and global network properties in order to convey important social phenomena and processes. Applications are used to illustrate the applicability of the presented approach, including when analysing the gendered inequalities in popular cinema by using dialogue networks. Moreover, some examples are provided on how to use the package ‘multigraphr’ in the statistical software R to perform the analysis.