Computational Communication Research - Volume 5, Issue 1, 2023

Spreading uncivil negative campaign messages is a “high-risk, high reward” campaign strategy since certain voters are more likely to be swayed by negative messaging whereas other voters are more inclined to feel sympathy with the attacked. Due to its risks, campaigns may attempt to outsource their uncivil ads to outside groups thus distancing themselves from the negativity and potentially avoiding any backlash. But at a time when advertising platforms boast of their ability to deliver ads to highly targeted audiences, uncivil negative ads could also be optimized to narrowly target citizens to which they are more likely to appeal. To study whether such optimizations are occurring, we retrieve all online advertisements that were placed on Facebook platforms (incl. Instagram) in the seven months prior to the US 2020 election. We perform multilevel ordinal regressions and find that ads from official political campaigns are more likely to be toxic when targeted at a narrower audience, whereas “dark money” outside groups (like super PACs and non-profits) are more likely to target broad audiences with their toxicity. In addition, we find that ads from outside groups are more likely to be toxic. We discuss the findings in light of this evidence and reflect upon future research regarding microtargeting negative messages on online platforms such as Facebook and Instagram.

Dictionary-based methods remain valuable to measure concepts based on texts, though supervised machine learning has been widely used in much recent communication research. The present study proposes a semi-automatic and easily implemented method to build and enrich dictionaries based on word embeddings. As an example, we create a dictionary of political incivility that contains vulgarity and name-calling words in Cantonese. The study shows that dictionary-based classification outperforms supervised machine learning methods, including deep neural network models. Furthermore, a small number of random seed words can generate a highly accurate dictionary. However, the uncivil content detected is only weakly correlated with uncivil perceptions, as we demonstrate in a population-based survey experiment. The strengths and limitations of dictionary-based methods are discussed.

In this work, we leverage a panel of over 1.6 million Twitter users matched with public voter records to assess how a standard keyword-based approach to social media data collection performs in the context of participatory politics, and we critically examine the speech this method leaves behind. We find that keyword classifiers undercount young people’s participation in online political discourse, and that valuable political expression is lost in the process. We argue that a mainstream keyword approach to collecting social media data is not well-suited to the participatory politics associated with young people and may reinforce a false perception of youth political apathy as a result.

Our study investigates the impact of social media on survey outcomes, particularly focusing on how the general public’s perception of issue importance is influenced by Twitter. To accomplish this, we propose two case studies in Swiss politics focusing on social media’s effect on opinion change during elections and the similarities in arguments on social media and in survey data during direct democracy votes. This study relies on a two-year longitudinal data collection of tweets (from January 2019 to December 2020) sent by more than 50.000 identified Swiss seed users and their most active followers. We find that the content of tweets surrounding popular votes reflects similar arguments as those expressed by a representative survey sample. Cases without congruence can be explained by the groups of users dominating social media discussions about the votes but also by the complexity and variety of arguments surrounding the votes. Furthermore, our findings demonstrate that the salience of policy issues on social media has no systematic effect on public opinion change regarding the perceived importance of similar issues, as social media’s impact strongly depends on the content and on the tweet’s author. These results provide valuable insights for researchers, helping them understand the circumstances in which social media data are most likely to be available and present a useful complement to surveys for understanding the dynamics of public opinion.

Topic Modeling is currently one of the most widely employed unsupervised text-as-data techniques in the field of communication science. While researchers increasingly recognize the importance of validating topic models and given the prevalence of discussions of inadequate validation practices in the literature, there is limited understanding of the consequences of employing different validation strategies when evaluating topic models. This study applies two different methods for topic modeling to the same text corpus. It uses four validation strategies to assess how the choice of validation method affects the final model selection and evaluation. Our findings indicate that different approaches and methods lead to different model choices and evaluations, which is problematic. This might lead to unwanted results in case the choice of model has a decisive impact on findings and, consequently, on theory development and practical implications.

One in four German internet users claims that search engines are their main gateway to news and a majority of Germans reports to primarily use their smartphone over their laptop/desktop computer to access news online. Yet, search-engine providers such as Google have repeatedly pointed out to actively favor specific forms of technical content optimization for mobile devices (e.g., Accelerated Mobile Pages), raising the question of whether this preference results in biases toward mobile-optimized content. In light of regulatory changes, this study investigates source diversity and source prominence in news items related to a range of issues presented to users when searching for news-related subjects via a smartphone or laptop/desktop computer in Germany. Using agent-based testing, 75,767 searches were performed on Google in mid-2020, applying a dynamic set of search terms on a range of different topics. Results indicate significant discrepancies in source diversity and source prominence between smartphone and laptop/desktop computers that can largely be attributed to mobile-optimized content likely to reach primarily younger news consumers who favor mobile devices. Overall, however, source diversity is equally high on both types of devices, highlighting the necessity for future research to focus on algorithmic influences on societies shared understandings of relevance beyond source diversity.

Media research is, in part, interested in accurately explaining and predicting people’s media selection. Explanation is an accurate description of the causal mechanisms that govern media selection whereas prediction is focused on making accurate inferences about unobserved data. However, meta-analyses demonstrate that existing media selection theories and models have limited explanatory accuracy. The predictive accuracy of these theories and models is unknown. Our project bridges this gap by empirically specifying how predictable, in principle, media selection is. To achieve this ambition, we articulate key conceptual distinctions between explanation and prediction. Subsequently, we report three empirical studies that examine prediction accuracy as a function of model complexity and estimate the theoretical maximum predictability of people’s music-listening and web-browsing behaviors. Approximately 80% of music selection and 60% of web-browsing behaviors are predictable. Moreover, a simple Markov Chain model that uses information about people’s prior media selection can achieve about 20% prediction accuracy for music selection and 10% accuracy in predicting web- browsing. By estimating the maximum predictability of people’s media selection behavior, we gain a first-ever benchmark by which media selection theories and models can be compared. More broadly, we show how simple models that account for the sequential dependency in media selection lend new insights and suggest novel directions for future theory development.

Profile photos, which are fundamental constructs of online self-presentation, are crucial to building successful online relationships. The visual cues embodied in profile photos are factors that help initiate socialization on visual-based gay dating platforms. This study applies app crawler techniques to collect publicly disclosed profile photos from Finka, a Chinese gay men’s dating app, and uses computer vision artificial intelligence applications to analyze the visual cues from the photos. The results show that appealing faces are relevant factors of user popularity, seemingly confirming the existence of the gay male beauty myth in the dating community. Photos that expose bodies are also popular, although most profile photos have low body visibility, which may be rooted in gay men’s culture in China. Extending existing research, we incorporate the background scene as another theoretical dimension of visual cues. Interestingly, we find an ambiance effect in profile photos—showing background scenes can serve as a self-presentation tactic that potentially influences users’ popularity. In particular, private scenes (e.g., locker room and bedroom scenes) can arouse more desire for further interaction with the presenter.

This paper introduces grafzahl, an R package for fine-tuning Transformers for text data from within R. The package is used in this paper to reproduce the analyses in other papers. Very significant improvement in model accuracy over traditional machine learning approaches such as Convolutional Neural Network is observed.

The automated content analysis of text has become integral to contemporary communication and journalism research. However, automated approaches are seldom utilized to analyze reported voice in text, while doing so would offer valuable insights into media and communication practices. Bridging the fields of communication science and computational linguistics, this study reviews and evaluates off-the-shelf tools for automated voice detection (of direct/indirect speech and of speakers) with respect to user experience and validity. Manually annotated English news articles and Twitter data served as baseline for evaluating the automated detection of voice. Findings indicate that the tools being assessed offer a satisfactory user experience and provide promising solutions for detecting direct speech automatically, encouraging fellow researchers to utilize automated detection for direct quotations. However, the recognition of indirect speech and speakers needs considerable improvement.

Counter-messages are used by civil education, youth prevention actors, and security agencies to counter the magnitude of polluted content. On the Internet, algorithmic operations of intermediaries affect how users encounter and receive polluted content. As counter-messages often show similar keywords, algorithms establish connections between counter-messages and polluted content, primarily because they share mutual topics. Against the background of legislative attempts to stop the spread of extremist online content, this paper aims to further investigate the interrelatedness of counter-messages and polluted content on YouTube due to the platform’s recommendation algorithm. To that end, two information network analyses were conducted based on each five seed videos of two differently designed counter-message campaigns one year after their publication on YouTube in 2019. Five thousand four hundred of the 35,982 videos of the two networks were analyzed qualitatively and manually. Results show that counter-messages are indirectly strongly connected to more polluted content. We further identify the campaigns’ design and setup on YouTube as factors that can cause the interrelatedness between counter-messages and polluted content.

Hurricanes are ungendered phenomena that are ascribed with gendered names. We examined if news information about the hurricanes are presented using gendered language. This work helps identify if people use gender stereotyping when referring to gender-neutral entities, and what these stereotypes might be. We use methods from natural language processing, qualitative text analysis, and statistics to analyze how gender is expressed in disaster-related news via text-level indicators: (1) pronouns, (2) lexical, syntactic, and semantic features of words related to hurricanes, and (3) types of sources quoted. Our sample contains news articles on 47 hurricane events from 1979 to 2012 from two weeks before to two weeks after landfall. We find that: (1) hurricanes are mainly referred to by gender-neutral pronouns, however, (2) when gendered pronouns are used, female-named hurricanes are five times more likely to be referred to by a gendered pronoun than male-named hurricanes, (3) adjectives and verbs used in discussing female-named hurricanes are on average more negative than those used for reporting on male-named-hurricanes, and (4) governmental sources are most frequently quoted as authority voices (voices from citizens and non-governmental entities are catching up), and a majority of these voices do not directly mention hurricanes with gendered references.

Taking into account YouTube’s specific role in the Russian media system and the increasing level of political polarization in the country, this study examines the role of incivility in discussions and whether discussions in an anti-government community represent a place for disagreement between pro-opposition and pro-government users. I argue that an online environment helps these sides meet each other rather than creating echo chambers of like-minded users. Moreover, in the quite restrictive Russian context for political deliberation, the incivility of messages plays a role in further involving commenters in discussions. Using the corpus of comments posted in the discussion section of opposition leader Alexei Navalny’s YouTube channel, I exploited class affinity modeling to identify pro-government and pro-opposition stances. Incivility was studied based on Google’s Perspective API toxicity classifier. I found that users avoid extreme forms of incivility when interacting with other commenters, but uncivil comments are more likely to start discussion threads. Furthermore, the level of incivility in comments gets higher over time after a video release. Pro- government sentiments, on the one hand, are associated with a subsequent response from Navalny’s supporters to the out- group criticism and, on the other hand, contribute to the further formation of hubs with a pro-government narrative. This research contributes to the extant literature on affective polarization on social media, shedding light on political discussions within an oppositional community in a non-democracy.