What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse

Shijia Zhou*, Siyao Peng*, Simon M. Luebke, Jörg Haßler, Mario Haim, Saif M. Mohammad, Barbara Plank · 2025

Read What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse by Shijia Zhou*, Siyao Peng*, Simon M. Luebke, Jörg Haßler, Mario Haim, Saif M. Mohammad, Barbara Plank (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-16592.txt.

Introduction

Shijia Zhou and co-authors build the first data set of climate memes labeled for both stance and framing, in a 2025 paper. Memes argue by image and caption, and how a meme frames climate change often betrays its author's stance. That link between frame and stance was largely unstudied. The team collected over a thousand memes from dozens of Reddit communities and annotated them using communication science categories.

The data reveal which frames each camp prefers and how tastes shift over time. Two machine tasks follow: detect a meme's stance, and detect its frame. Vision-language models did well on stance but stumbled on frames, where plain language models did better. Adding human-written captions reliably helped the machines read the joke.

Catalog record

CREATORShijia Zhou*, Siyao Peng*, Simon M. Luebke, Jörg Haßler, Mario Haim, Saif M. Mohammad, Barbara Plank
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-16592

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse Shijia Zhou*, Siyao Peng*, Simon M. Luebke, Jörg Haßler, Mario Haim, Saif M. Mohammad, Barbara Plank Abstract Media framing refers to the emphasis on specific aspects of perceived reality to shape how an issue is defined and understood. Its primary purpose is to shape public perceptions, often in alignment with the authors’ opinions and stances. However, the interaction between stance and media frame remains largely unexplored. In this work, we apply an interdisciplinary approach to conceptualize and computationally explore this interaction with internet memes on climate change. We curate ClimateMemes , the first dataset of climate-change memes annotated with both stance and media frames, inspired by research in communication science. ClimateMemes includes 1,184 memes sourced from 47 subreddits, enabling analysis of frame prominence over time and communities, and sheds light on the framing preferences of different stance holders. We propose two meme understanding tasks: stance detection and media frame detection. In various input setups, we evaluate two vision language models (VLMs), LLaVA-NeXT and Molmo, and report the corresponding results on their backbone large language models (LLMs). Human captions consistently enhance performance. Synthetic captions and human-corrected OCR also help occasionally. Our findings highlight that VLMs perform well on stance, but struggle on frames, where LLMs outperform VLMs. Finally, we draw on concepts from Computational Communication Science to analyze VLMs’ limitations, showing that memes employing specific humor types, personalization, and responsibility cues pose challenges for VLMs in handling nuanced frames and stance expressions on climate change. 1 1 1 We make our dataset publicly available at https://github.com/mainlp/ClimateMemes . 1 Introduction (a) convinced stance with real and impact frames (b) skeptical stance with hoax frame Figure 1: Two climate change memes conveying opposite stances using different media frames. Internet memes are a powerful communication format in online discourse that reflect communities’ cultural and social dynamics Davis et al. (2016) ; Zhang and Pinto (2021) . Multimodal digital items combine images and texts to convey complex viewpoints in a compact and engaging format Sharma et al. (2020) ; Liu et al. (2022) . Through these multimodal expressions, communicators convey their positions towards particular topics, i.e., stances as defined by Mohammad et al. (2016) . While stance reflects the creator’s opinion toward a target, the specific narrative used to convey a certain stance is shaped by media frames. Media framing refers to selecting specific aspects of a perceived reality in communication to portray how an issue is defined, how its causes are interpreted, how its moral implications are evaluated, and what potential solutions are considered Entman (1993) ; Gidin (1980) . Depending on their stance, creators may gravitate toward different framing strategies Snow and Benford (1992) . However, the interaction between stance and media frames remains under-studied, particularly in their representation through humorous social media content such as memes. This is especially relevant for debates of global significance, such as climate change. Memes about climate change (CC) are widespread on social media, including Twitter/X Ross and Rivers (2019) . For example, Figure 1(a) conveys a convinced stance towards CC by using real and impact frames (further detailed in § 3.2 ) to affirm the evidence of global warming and its disheartening consequences. Conversely, Figure 1(b) conveys a skeptical stance using the hoax frame, claiming that CC is not a major issue or even not real, and suggests that politics may distort the CC issue. In this paper, we analyze stances and media frames in CC memes by examining the following three research questions (RQs): • RQ1: How do different media frames shape the visual representation of climate change in memes across varying stances? We introduce ClimateMemes , a dataset of CC memes, consisting of 1,184 CC memes from 47 subreddits, manually annotated with stance on climate change and the media frames they invoke (§ 3 ) to analyze how memes convey stance through strategic media framing (§ 4 ). • RQ2: Can state-of-the-art VLMs and LLMs accurately detect stances conveyed by memes and the corresponding media frames? We extend stance detection from text and propose a new task of multi-label media frame detection on CC memes. We evaluate two open-source VLMs and their backbone LLMs (§ 5.1 ) and investigate the effects of few-shot experiments and input modalities on these two tasks (§ 5.2 ). We found that while synthetic meme captions cannot yet fully replace human-annotated ones, they still improve the VLMs’ performance on both tasks. Yet, LLMs outperform VLMs on frame detection. • RQ3: Can taxonomies from communication science provide more insights into stance and media frame detection results? We recruit communication science specialists to annotate humor type, person, and responsibility features on 235 test CC memes. Our analyses reveal that the performances of VLMs and LLMs degrade markedly on memes that are jokes, about political figures, and about individual (micro-level) responsibilities (§ 6 ). 2 Background Figure 2: ClimateMemes ’s pipeline of data collection, filtering, and annotations of stance, media frames, etc. 2.1 Memes Internet memes are multimodal and humorous forms of expression that are popular across various digital channels, especially on social media Shifman (2014) . They often use replicated and modified templates and are circulated among users to convey new, context-specific meanings. For example, Figure 1(a) uses the “Simpsons so far” template to highlight the ongoing trend of global warming. In controversial political arenas, such as the climate discourse, internet memes are seen as an effective tool for capturing attention, allowing users to communicate their stances through impactful imagery and humor Ross and Rivers (2019) . Nguyen and Ng (2024) categorize meme understanding into three main types: classification , interpretation , and explanation . Classification aims to assign labels to memes, such as identifying harmful content Kiela et al. (2020) ; Pramanick et al. (2021) ; Cao et al. (2022) ; Hee et al. (2023) ; Chen et al. (2025) ; Liu et al. (2025) ; Chakravarthi et al. (2025) , sentiment Sharma et al. (2020) ; Chauhan et al. (2020) ; Nguyen et al. (2025) , or figurative language Liu et al. (2022) ; Xu et al. (2022) . Interpretation tasks focus on understanding and generating insights from memes, such as generating captions or analyzing the metaphor between the image and text componants Hwang and Shwartz (2023) ; Chen et al. (2024) . Explanation tasks go a step further by generating textual justifications for the labels assigned to memes ( Hee et al., 2023 ) . In this study, we curate ClimateMemes and introduce two meme understanding tasks: stance detection and media frame detection. We also collect human-corrected OCR and human-written meme captioning as a basis for future tasks. 2.2 Media Frames Strategic media framing refers to the selective presentation of information to influence audience attitudes or evoke specific reactions Snow and Benford (1992) . Social and communication science research has relied on framing concepts for analyzing how information is selected and presented in the media. Scholars in the field have created codebooks for manual identification of generic and issue-specific frames in media contexts. The Media Frames Corpus ( Card et al., 2015 ) , focusing on three specific issues: immigration, smoking, and

Read the complete text