Content Analysis of Mediated Associations: An Automated Text-Analytic Approach

ABSTRACT

Due to the fact that mediated associations are a central aspect of many mass communication theories, their measurement is of central interest for communication research. Mediated associations are defined as the repeated pairing of an object (e.g., social group, political party) with specific attributes (e.g., crime, economy). In this article, we introduce a recently developed, automated text-analytic technique. We present an application of this method in the media stereotyping domain via the content analysis of German news coverage of Islam. As predicted, the analysis revealed substantial mediated associations between Islam-related concepts and violence (e.g., “Koran + violence”), terror (e.g., “Islam + terror”), dehumanization (e.g., “Muslims + animal-related terms”), and general negativity (valence). We discuss the promises and pitfalls of this method, make software suggestions, and provide application-related information for speedy dissemination in communication research.

Notes

^1. New Year’s Eve celebrations in Cologne were disrupted by robbery, theft, and sexual offenses committed by a crowd of young men mostly of North African and Arab descent. During a street party at Central Station, hundreds of young men formed rings around young women. While some of these women were robbed, others were sexually harassed. Most of these men seemed to be drunk, aggressive, and without inhibitions, as both the police and witnesses have reported (Report of the Federal Ministry of the Interior of North Rhine-Westphalia, 2016). After this incident, the Cologne police received a total of 1,054 complaints, with 454 being related to sexual assault (Flade et al., 2016).

^2. These techniques have also been used to build associative networks in order to represent the structure of text (Diesner & Carley, 2005). Although network analyses based on these techniques are named differently (map analysis: Diesner & Carley, 2004; semantic mapping: Leydesdorff & Welbers, 2011; semantic network analysis: Doerfel & Barnett, 1999), these approaches consist of concepts and their linkages (Carley, 1997). The idea is that concepts that frequently co-occur are more strongly related and therefore closer to each other than those that rarely co-occur (Rosen, Woelfel, Krikorian, & Barnett, 2003). This is similar to a regular CASS analysis. However, CASS is intended to be used merely in a deductive way: Researchers use specific target concepts that have been defined prior to analysis. This deductive approach also distinguishes CASS from other sematic space models such as LSA (Landauer & Dumais, 1997) which reduces a corpus to fewer latent dimensions with each dimension containing words that are semantically related (Sahlgren, 2006). Hence, LSA only relies on statistical analyses and therefore is—in contrast to CASS—free from a theory-driven coding scheme (Simon & Xenos, 2004). If a scholar’s goal is to use automated text-analytic approaches to reveal entire associative networks (i.e., a more inductive approach), then other approaches to automated content analysis (e.g., semantic network analysis) are presumably more appropriate.

^3. Measuring the association of predefined targets words rather than computing the association of the most frequently occurring words in a co-occurrence analysis usually requires the application of word lists, since not all concepts are described by a single word but rather by different synonyms. Therefore, it should be noted that using word lists (e.g., BAWL) as described above corresponds to some procedures of the dictionary approach (van Atteveldt, 2008).

^4. The CASS-software (Holtzman et al., 2014) allows for the calculation of a mean Δz-score over all items. In our case, the five Islam/Christianity-related word pairs. This score represents the “average” mediated association. Importantly, the computation of a mean Δz-score only makes sense when the items show high internal consistency. Due to the fact that we defined the word pairs a priori to data analysis, we refrain from reporting a mean Δz-score and the corresponding tests of significance to ensure a strictly deductive approach.