(iPoster) Identifying Visual Power: How Nationalist Visual Elements Resonate with Netizens

• In Event: Poster Sessions: Political Communication II & The Political Power of Film and Reality TV
  In Poster Session: Poster Session: Political Communication II

Abstract

Scholars have increasingly recognized the important role of visual content in unpacking media effects, particularly in the context of political propaganda. A growing body of research consistently documents a seemingly intuitive yet empirically robust finding: social media posts featuring visual content tend to generate significantly higher audience engagement compared to their textual counterparts. However, prior studies have largely overlooked a crucial confounding factor—the heterogeneity of textual content embedded in media posts, which may lead to an overestimation of the causal influence of visual elements in engaging audiences. A plausible alternative explanation is that posts accompanied by visual attachments might inherently feature more compelling or catchy textual content. While a set of studies have acknowledged these confounding effects and attempted to mitigate them, their approaches have predominantly relied on occurrence-based matching techniques, such as n-gram analysis or more sophisticated Bag-of-Words (BoW) models, or superficial semantic proxies like the latent Dirichlet allocation (LDA) models. These methods fail to capture the full semantic richness of textual content, thereby limiting their ability to accurately estimate the impact of visual elements. In this study, we propose an approach that efficiently matches textual data at the semantic and contextual levels and identifies target visual elements rather than non-representative features, such as tone, brightness, or layout. Specifically, we first utilize the state-of-the-art text-to-vector methods of large language models (LLMs) to convert text into high-dimensional vectors embedded with rich contextual and semantic information. The semantically similar posts are matched by measuring the similarity between texts based on vector similarity. We then deploy vision language models (VLMs) with advanced inferential capabilities to label the extracted target visual features. This set of semantically matched and visually labeled posts, in turn, enables researchers to isolate and estimate the real effects of visual content. We apply our methodological framework to a dataset of over 10,000 Weibo posts from influential official accounts regarding the 2019 Hong Kong social movement and ask: Are nationalist visual elements more effective in evoking audience resonance on Chinese social media in this context? We begin by demonstrating that semantically matched posts display no statistically significant differences in audience engagement levels. This finding establishes a solid methodological basis for isolating and estimating the causal effects of visual elements. Furthermore, in a departure from the prevailing consensus in existing literature that visual elements consistently yield substantial gains in audience engagement, our study reveals a curvilinear association between visual elements and nationalist engagement, a pattern that persists regardless of whether the textual content is matched or not. Specifically, we find that, before semantically matching posts, extreme nationalist elements are statistically significant and positively associated with higher levels of engagement, whereas mild or tempered nationalist elements tend to resonate less with the digital audience. However, we find it striking that, after matching textual content, extreme nationalist visual elements do not statistically outperform their textual counterparts in terms of driving audience engagement. In doing so, our contribution is twofold: first, scholars should rigorously account for the confounding influence of narrative strategies when assessing the effects of visual elements; second, it provides a viable analytical framework for more accurately quantifying such visual effects.

Author

• Unregistered Participant