Computational Communication Research - Volume 8, Nummer 1, 2026

This paper presents a novel approach to classifying news articles by topic using only their URLs, addressing growing challenges in accessing article text due to paywalls and scraping restrictions. By fine-tuning a DistilBERT transformer model on URL data alone, I demonstrate topic classification performance that matches or exceeds traditional approaches requiring article text. Across three benchmark datasets spanning multiple languages and over 660,000 articles from more than 11,000 news domains, this URL-based topic classifier achieved superior F1 scores compared to both conventional machine learning methods and existing URL-based techniques. While this method requires more computational resources than simpler topic classification approaches, it dramatically reduces data collection requirements, offering researchers a practical alternative when text access is limited. These findings suggest that news article URLs contain richer semantic information than previously recognized, opening new possibilities for large-scale news content analysis in increasingly restrictive digital environments.

Computational methods can minimize the time and resources needed to manually code thousands of images. Yet, they also come with challenges, including validation, algorithmic bias, and privacy concerns. Acknowledging that the pictorial turn has now entered a computational phase, this article reports on a manual and automated coding of 7000+ images to better understand online extremist content. Using Rodriguez and Dimitrova’s (2011) four-tiered model of visual framing, the study compares manual and OpenAI’s ChatGpt4o’s coding of Al-Qaeda and ISIS images across the denotative, semiotic, connotative, and ideological levels. AI coding exhibited moderate to strong performance on denotative variables but was weaker in the semiotic and connotative tiers. The study concludes with a discussion of the advantages of human and AI functioning together to better understand visual framing.

This research article presents insights from a two-wave, longitudinal data donation study across four major social media platforms: TikTok, YouTube, Facebook, and Instagram. We investigate a critical yet underexplored aspect of data donation: allowing participants to delete specific data traces before submission. Our analysis quantifies the impact of this selective omission on data completeness and, consequently, the analytical power of the resulting datasets. Furthermore, leveraging a longitudinal design, we examine the stability of donation and deletion behaviors over time in a panel setting. Findings reveal an overall increase in the platform donor rate in the second wave. However, we also observe substantial donor attrition. Notably, the omission of data traces is predominantly observed among first-time donors.Our results suggest the feasibility of longitudinal data donation research designs. For allowing participants selective data omission, a careful weighing of the trade-offs is necessary, as this practice—when utilized—significantly compromises data completeness.

This paper leverages large-language models (LLMs) to experimentally determine strategies for scaling up social media annotation and stance detection of health information, with HPV vaccine-related tweets as a case study. We examine both conventional fine-tuning and emergent in-context learning methods, systematically varying strategies of prompt engineering and in-context learning across widely used LLMs and their variants (e.g., GPT-4, Mistral, Llama 3, and Flan- UL2). Specifically, we varied prompt template design, shot sampling methods, and shot quantity to detect stance on HPV vaccination. Our findings reveal that (a) in-context learning outperformed fine-tuning in stance detection for HPV vaccine social media content; (b) increasing shot quantity does not necessarily enhance performance across models; (c) stratified sampling often outperforms random sampling, with the performance gap more pronounced in smaller model variants, and (d) LLMs and their variants present differing sensitivity to in-context learning conditions. This study highlights the potential and provides an applicable approach for applying LLMs to research on social media annotation and stance detection of health information.