The emergence of social media has led memes to become a powerful mode of communication, blending text, images, and emojis. However, this surge in meme usage has also seen a rise in offensive material. With manual content moderation proving impractical due to the sheer volume of data, there's a pressing need for automated methods to identify harmful memes. Yet, existing research predominantly targets high-resource languages such as English, neglecting low-resource ones like Nepali. To bridge this gap, we introduce the first Nepali meme dataset annotated for hate speech and sentiment. Our contributions are threefold: (1) We create and release NeMeme, a unique dataset featuring Nepali and code-mixed Nepali memes (combining Nepali and English). (2) We evaluate NeMeme using cutting-edge unimodal and multimodal models to establish initial performance benchmarks. (3) We introduce MemeNePAL, a novel multimodal framework employing prompt-assisted learning to effectively categorize Nepali memes. MemeNePAL overcomes the shortcomings of prior state-of-the-art (SOTA) techniques, which were designed for high-resource languages and struggle with Nepali's linguistic differences and cultural subtleties. This work not only promotes inclusivity in content moderation research but also aligns with UN Sustainable Development Goals such as promoting well-being, reducing inequalities, and fostering peace. We adhere to FAIR principles by making the dataset publicly available.
This study examines how individual music listening behaviors evolved during the COVID-19 lockdowns in France, focusing on both listening volumes and rhythms. We combine passively collected individual listening history data, provided by a music streaming service and covering the 2019-2023 period, with survey data collected from the same users (n ≈ 10000). Using the Dynamic Time Warping method, we develop a typology of listening trajectories during the first lockdown. The results reveal significant and heterogeneous changes in listening behavior, with approximately one-third of respondents experiencing a significant decrease in listening volume, while a quarter experienced an increase. We then analyze the evolution of the intervals between consecutive music listening sessions — so called inter-session times — to assess disruptions in individual listening rhythms. We uncover an unprecedented shift in the listening rhythms at the onset of first lockdown, reflecting varying degrees of disruption in daily life rhythms. For half of the individuals this disruption lasted more than four weeks. Finally we show that age, educational attainment and household structure unevenly influence the reorganization of music listening activity during this period, shedding light on the social differentiations at work in the reorganization of an ordinary activity during this crisis period.
Ensuring the online safety of youth has motivated research towards the development of machine learning (ML) methods capable of accurately detecting social media risks after-the-fact. However, for these detection models to be effective, they must proactively identify high-risk scenarios (e.g., sexual solicitations, cyberbullying) to mitigate harm. This `real-time' responsiveness is a recognized challenge within the risk detection literature. Therefore, this paper presents a novel two-level framework that first uses reinforcement learning to identify conversation stop points to prioritize messages for evaluation. Then, we optimize state-of-the-art deep learning models to accurately categorize risk priority (low, high). We apply this framework to a time-based simulation using a rich dataset of 23K private conversations with over 7 million messages donated by 194 youth (ages 13-21). We conducted an experiment comparing our new approach to a traditional conversation-level baseline. We found that the timeliness of conversations significantly improved from over 2 hours to approximately 16 minutes with only a slight reduction in accuracy (0.88 to 0.84). This study advances real-time detection approaches for social media data and provides a benchmark for future training reinforcement learning that prioritizes the timeliness of classifying high-risk conversations.
This study delves into the mechanisms that spark user curiosity driving active engagement within public Telegram groups. By analyzing approximately 6 million messages from 29,196 users across 409 groups, we identify and quantify the key factors that stimulate users to actively participate (i.e., send messages) in group discussions. These factors include social influence, novelty, complexity, uncertainty, and conflict, all measured through metrics derived from message sequences and user participation over time. After clustering the messages, we apply explainability techniques to assign meaningful labels to the clusters. This approach uncovers macro categories representing distinct curiosity stimulation profiles, each characterized by a unique combination of various stimuli. Social influence from peers and influencers drives engagement for some users, while for others, rare media types or a diverse range of senders and media sparks curiosity. Analyzing patterns, we found that user curiosity stimuli are mostly stable, but, as the time between the initial message increases, curiosity occasionally shifts. A graph-based analysis of influence networks reveals that users motivated by direct social influence tend to occupy more peripheral positions, while those who are not stimulated by any specific factors are often more central, potentially acting as initiators and conversation catalysts. These findings contribute to understanding information dissemination and spread processes on social media networks, potentially contributing to more effective communication strategies.
Manually annotating data for computational social science tasks can be costly, time-consuming, and emotionally draining. While recent work suggests that LLMs can perform such annotation tasks in zero-shot settings, little is known about how prompt design impacts LLMs' compliance and accuracy. We conduct a large-scale multi-prompt experiment to test how model selection (GPT-4o, GPT-3.5, PaLM2, and Falcon7b) and prompt design features (definition inclusion, output type, explanation, and prompt length) impact the compliance and accuracy of LLM-generated annotations on four highly relevant and diverse CSS tasks (toxicity, sentiment, rumor stance, and news frames). Our results show that LLM compliance and accuracy are prompt-dependent. For instance, prompting for numerical scores instead of labels reduces all LLMs' compliance and accuracy. Concise prompts can significantly reduce prompting costs but also lead to lower accuracy on tasks like toxicity. Furthermore, minor prompt changes like asking for an explanation can cause large changes in the distribution of LLM-generated labels. By assessing the impact of prompt design on the quality and distribution of LLM-generated annotations, this work serves as both a practical guide and a warning for using LLMs in CSS research.
