Rules and monitoring can help communities coordinate under stress, but they can also harden into counterproductive red tape. This paper asks whether a wiki’s pre-shock procedural regime shapes how it reorganizes when stress spikes. Bureaucratic state is measured as excess monitoring (RexSOC)—monitoring activity above what would be expected given scale, workload, and conflict within each wiki-month. Using a monthly panel of 82 Wikimedia wikis (2004–2026), local-projection event studies are estimated around extreme conflict months, with overload months used as a secondary exploratory comparison. The main outcome is enforcement concentration, measured as the top-1
This study analyses the amplification and diffusion of climate misinformation on Twitter during the COP26 and COP27 climate conferences. Drawing on a dataset of over 12 million English-language tweets, we combine machine learning classification, social network analysis, and qualitative content analysis to map how misinformation circulates across user communities. Climate misinformation is understood as content that denies or undermines the scientific consensus on anthropogenic climate change. Using a machine learning classifier trained on annotated climate datasets, tweets were labelled and assigned misinformation probabilities. Using community detection, we were able to distinguish between misinformation, non-misinformation, and mixed communities. Our findings show that misinformation does not remain isolated within echo chambers; instead, it often flows outward, particularly into mixed communities, which serve as key intermediaries between misinformation and non-misinformation communities. Through centrality measures, we identified a small set of influential user accounts that function as amplifiers and brokers of misinformation, both intentionally and inadvertently. These key users exhibit varying patterns of visibility, engagement, and connectivity, where we found two prominent user types to be those of the broadcaster and the mediator. The results show that the dynamics of misinformation dissemination are shaped by both content virality and underlying network structures.
This study investigates the Italian odonymic landscape through the lens of critical toponymy, employing computational methods for semantic classification to examine gendered patterns of commemoration in street naming. A national-scale database is constructed by integrating official toponymic registers with rule-based heuristic procedures and large language model-based classification. This hybrid infrastructure enables the systematic identification and categorization of odonyms dedicated to individual and collective persons, with particular attention to gender representation. The combined classification approach reliably detects personal commemorations and infers gender attributes, yielding results that closely align with manually curated datasets. The study also advances critical toponymy by showing how computational methods can operationalize analyses of collective memory and symbolic power in urban space. The results reveal the persistent gender asymmetries in Italian commemorative odonymy.
Advances in machine learning have enhanced researchers’ ability to detect vaccine hesitancy on social media using Natural Language Processing. Our objective in this study was to evaluate, through a systematic review of studies that employed machine learning to analyze sentiment and stance regarding COVID-19 vaccines on Twitter in assessing their methodological quality and consistency. We searched for papers published between 1 January 2020 and 31 December 2023 in PubMed, Web of Science, and Scopus. The inclusion criteria were the use of supervised machine learning to assess COVID-19 vaccine hesitancy through stance detection or sentiment analysis on Twitter/X. We categorized the studies according to a taxonomy of five dimensions: tweet sample selection approach, self-reported study type, classification typology, annotation codebook definitions, and interpretation of results. We analyzed the risk of bias in the included studies, examining whether stance detection was used to report different hesitancy trends compared to those using sentiment analysis. We identified 51 papers that were published in 36 journals. Our review found that measurement bias is widely prevalent in studies employing supervised machine learning to analyze sentiment and stance toward COVID-19 vaccines and vaccination. The reporting errors are sufficiently serious to hinder the generalizability and interpretation of these studies, making it difficult to determine whether discursive content communicates reluctance to vaccinate against SARS-CoV-2. Our findings underscore the importance of more transparent reporting of NLP methods in vaccine discourse studies. Addressing methodological shortcomings is essential to improving our understanding of vaccine hesitancy on social media.
This study investigates whether machine learning models trained on social media data can reliably predict real-world Black Lives Matter (BLM) protests in New York City, with a focus on the importance of temporal dynamics in the online-offline relationship. Using the Elephrame protest database and Giorgi's Twitter Corpus covering August 2014 to December 2020 (239 protests; 28,061 tweets), we apply temporal data splitting to prevent data leakage and compare feature engineering strategies across seven algorithms while addressing class imbalance. The findings reveal that simple models with well-chosen features can reliably predict protest events. Notably, hashtag volume dynamics are significantly more predictive than semantic content, indicating that engagement metrics capture mobilization signals more effectively than text analysis. The results also demonstrate concept drift, showing that the digital signal patterns associated with protests evolve over time. Overall, the results show that social media discourse can effectively predict offline protests. They support collective action theory’s emphasis on mobilization and political process theory’s focus on planning capacity. However, concept drift suggests that future predictive systems must be dynamic and adaptive to remain effective.
Effective flood management in urban planning relies on accurate, timely data, which can be sourced from social media platforms for real-time post-flood damage information. However, many social media content lack location, creating a significant challenge for spatial analysis. This study addresses this gap by proposing a novel framework to infer the locations of post-flood events extracted from social media content, leveraging flood vulnerability maps and a structured knowledge base. The methodology involves four key steps, extracting flood-related events using hypergraph-based clustering; creating bounding boxes for potential event locations by integrating flood-related keywords, spatial proximity, and temporal patterns; constructing a knowledge base incorporating flood vulnerability criteria; and inferring event locations by comparing non-geo-tagged events against the knowledge base rules and aligning them with spatiotemporal bounding boxes. By analyzing 150,000 social media posts from flood-affected regions in southwestern Iran, such as Ahvaz, between April 6–16, 2019, the method identified 27 flood and 1200 post-flood events; of the 970 non-geo-tagged events, 69 were inferred inside the study region, while the remaining 901 were inferred out-of-region and were not mapped. Evaluation metrics, including 70
In an era of deepening political polarization, corporate political activism has emerged as a critical nexus of consumer-brand interaction. This study investigates the alignment between U.S. corporations’ offline political donations and the ideological composition of their online audiences on the X (formerly Twitter) platform. While existing research often relies on surveys, extensive empirical evidence from real-world social media remains limited. To address this gap, this research employs a novel computational approach, analyzing a large-scale X dataset to construct a political retweet network. Audience ideology is inferred using a validated community detection method. The analysis finds a robust association between the partisan direction of corporate donations and the ideological makeup of their X audience, a relationship that persists after controlling for firm-level factors. Furthermore, ideological misalignment is associated with a meaningful increase in negative online engagement. This study provides evidence consistent with social identity theory in consumer-brand dynamics and demonstrates the utility of computational methods for linking offline political actions to online social structures, offering crucial insights for navigating reputational risk in a polarized digital landscape.
Reliable estimation of child trafficking prevalence at fine geographic scales is critical for targeted program and policy interventions, yet such estimation is often hindered by sparse or incomplete data. This study investigated the use of small area estimation (SAE) methods—one frequentist and the other hierarchical Bayesian—to improve prevalence estimates using data from a household survey of 3070 households across three hotspot districts in Sierra Leone. SAE models were implemented at the chiefdom level (m = 40), the lowest administrative unit, and compared against traditional district-level estimates. Both SAE approaches yielded chiefdom-specific prevalence estimates and predictive probabilities that exhibit substantial heterogeneity and greater precision than the district-level estimates. By borrowing strength across small areas, SAE methods achieved notable reductions in variance and enhanced inference for rare and sensitive outcomes. These findings underscore the utility of SAE for producing reliable local-level estimates in contexts where direct estimation for small geographic areas is infeasible or insufficient. The study highlights the methodological and practical implications of SAE to inform localized interventions and encourages its broader application in research on hidden or hard-to-reach populations, including those affected by crime victimization and human rights violations.
In recent years, studies have revealed a decline in semantic variety across popular music lyrics, particularly in English-language songs on streaming platforms like Spotify. This research examines whether a similar trend can be observed in a different linguistic and cultural context: the lyrics of all finalist songs from the 75 editions of the Sanremo Music Festival, Italy’s most renowned music competition. What sets this work apart is the development of a flexible and efficient methodology for tracking changes in semantic similarity over time, which can be applied to different datasets to study similar phenomena. Drawing on a combination of full-text, segment-based, topic-based, and word-level analyses, the approach leverages both embedding techniques and large language models. When applied to the Sanremo corpus, this framework reveals a gradual move toward increasing semantic uniformity, echoing the global patterns identified in previous studies. These findings underscore the value of natural language processing tools in uncovering long-term shifts in musical language and cultural expression.
A model of interconnected socioeconomic and health status based on quantitative evidence and qualitative information was built during the pandemic in Indonesia. The integrated resilience framework was established, which is becoming the basis of analyzing the system's ability to withstand COVID-19 and drawing lessons learned after COVID-19. The system dynamics method was used by integrating quantitative evidence with qualitative information. The system dynamics simulation revealed that first, adaptive health management, which balanced public health concerns with socioeconomic considerations, and second, socioeconomic resilience were characterized by i. a wide range of social initiatives that arose to increase resilience to the social impact of COVID-19, and ii. economic resilience, which demonstrated the diversity of business creativity that emerged to survive the COVID economic impact. The model of integrated resilience as an interconnected system of people's socioeconomics and government-led adaptive health management was able to withstand COVID-19. The lessons learned from the system's ability to withstand COVID-19 are i. the emergence of adaptive health management through experiential learning and ii. people’s innovativeness increased because of digital technology adoption.
This paper introduces a computational framework for classifying entity role framing in unstructured textual data, with a significant application in quantitatively measuring and understanding political bias in news media. Traditional approaches to political bias analysis have often relied on less interpretable quantitative metrics, creating a demand for more transparent and nuanced computational solutions. The interpretability of our approach stems from the explicit identification of entities and the specific roles (hero, villain, victim) assigned, offering a direct insight into the narrative strategies employed. Building on previous works on computational modelling of entity framing in multimodal content, we examine the efficacy of textual features for this task, using three experimental setups of increasing complexity and two language models. Crucially, when applied to a corpus of politically diverse news outlets in the US and UK, the framework successfully quantifies polarised patterns in how media sources frame political entities, demonstrating clear alignment with established media bias measurements. This methodological framework provides a systematic approach to mapping entity framing patterns to political orientation, thereby creating a more interpretable and nuanced measure of polarisation in digital news ecosystems. The framework can serve as a valuable tool for scholars and practitioners in political communication, journalism, and computational social science seeking to understand the dynamics of digital news ecosystems.
The rising frequency of crimes against women necessitates the development of the accurate and effective methodology to enable targeted interventions and preventive measures. This research utilizes a dataset prepared from the National Crimes Record Bureau’s 2022 report on crimes against women. In this study, we investigated the application of advanced ensemble classification techniques to analyze the crime against women dataset. The spatial distribution of aggregated crimes is illustrated using a map of India, providing a visual representation of how crimes against women are distributed across the country. Our proposed approach employs a hybrid ensemble model that integrates the base learners as Light Gradient Boosting Machine, CatBoost, and AdaBoost, applied with the novel hybrid meta-learner. The novel hybrid meta-learner incorporates Random Forest and CatBoost as base learners, with logistic regression serving as meta-learner. By stacking these base learners and applying the novel hybrid meta-learner, this approach capitalizes on the strengths of each technique, overcoming their individual limitations and enhancing overall classification performance. We evaluate performance of baseline models, ensemble models, and our proposed hybrid ensemble model using metrics, accuracy, precision, recall, and F1-score. Our comparative analysis showed that the proposed hybrid model surpasses both baseline and conventional ensemble methods in classification accuracy and reliability. Additionally, the efficiency of the Meta_RCL model is validated through cross-validation, achieving an accuracy of 0.969, precision of 0.969, recall of 0.968, and F1-score of 0.968 which are the highest compared to other models. This model’s versatility allows it to be applied to other crime datasets, aiding law enforcement agencies, policymakers, and social organizations in identifying patterns, trends, and risk factors.
Hate speech on social media poses a serious threat to online safety and social harmony, especially in multilingual and low-resource contexts such as Indonesia. The challenge is further amplified by the increasing prevalence of multimodal contents that combine both text and images, which are often used to convey harmful messages more subtly. However, most existing research focuses only on text-based detection, leaving a gap in understanding how visual information contributes to hate speech utterance. This study aims to comprehensively investigate hate speech detection on Indonesian social media within a multimodal setting. We constructed a novel multimodal hate speech dataset that includes tweet text and accompanying images collected from X, with annotations provided by experts. For evaluation, we explore five hate speech detection approaches: unimodal baselines, early fusion, late fusion, multimodal large language models (MLLMs), and textualization-based fusion. Additionally, we also employed traditional machine learning and transformer-based models for more comprehensive assessment. Our experiment results show that textual information alone remains highly predictive, with a strong baseline F1-score of 0.803 achieved by XLM-RoBERTa. MLLMs such as Gemini 2.5 Flash and GPT−5.2 demonstrate competitive zero-shot performance (up to F1 = 0.799), though with variability across languages and prompting strategies. Our experiment also uncovered that combining raw image and text via early or late fusion yields only marginal gains. In contrast, our proposed textualization approach by transforming images into descriptive text using GPT-4o and concatenating them with tweet text achieves the best overall result with an F1-score of 0.833. These findings suggest that structured textual representations of visual content can significantly enhance multimodal classification performance while improving interpretability. This work provides valuable insights for future research on hate speech detection, especially in resource-scarce and linguistically diverse regions.
Detecting Depression in individuals can prevent significant amounts of mental anguish. This study expands on our previous work, “Detecting Depression: Employing Natural Language Processing and Random Forests”, and further explores the potential of machine learning in detecting Depression from text. Previously, we curated a dataset of 291 Depression and 698 Control text posts obtained from Reddit. Using TF-IDF features, we trained Random Forest classifiers and achieved an F1 score of 93.24
Identifying influential nodes in social media networks is critical for improving information dissemination, digital marketing, and crisis response strategies. Traditional centrality measures exhibit key limitations: local metrics are fast but overlook structural context, while global metrics reflect network-wide influence but are computationally expensive and insensitive to local diversity. Hybrid approaches attempt to balance both, yet often neglect how a node’s influence depends on the exclusivity and structural positioning of its connections. To address these limitations, we propose Influence Disparity Centrality (IDC), a novel centrality measure that evaluates node importance not only by connection quantity, but by relative degree superiority and neighborhood non-redundancy. IDC prioritizes nodes that serve as non-obvious bridges between weakly connected regions, enhancing global diffusion with minimal redundancy. We validate IDC across heterogeneous datasets including Facebook networks, academic and student social graphs, and thematic online communities. Quantitative evaluations, which include Pearson and Spearman correlations, Kendall’s Tau under SIR simulations, as well as Epidemic Duration and Monotonicity, demonstrate that IDC consistently outperforms both traditional and hybrid measures in identifying structurally and functionally influential nodes. Unlike classical metrics which saturate early or misidentify hubs, IDC yields broader and more sustainable propagation in dynamic scenarios. Its capacity to reveal structurally strategic actors makes it particularly useful for large-scale influence modeling, viral outreach, and online behavioral interventions.
The rapid dissemination of graphic and violent imagery in recent times has outpaced existing moderation tools, and yet most public benchmarks focus on only one type of inappropriate content and treat moderation as a simple binary task—Pornography vs. Neutral. This research introduces C4Censor, a lightweight, multi-class image benchmark for fine-grained censorship across four high-risk categories Blood Gore, Pornography, Terrorism, and Neutral—each subdivided into three challenging subclasses (e.g., Hentai vs. Anime, Counter-Terrorism vs. War-Crimes) with 500 images per subclass, for a total of 6k images. All visuals were scraped from publicly accessible sources—gaming streams (YouTube, Twitch), medical procedure archives, specialty NSFW (not safe for work) collections, X (formerly Twitter) and Telegram channels tracking extremist activity, and curated public websites—and each image was manually annotated at both the coarse (4-way) and fine (12-way) levels. Unlike existing datasets that are either binary or single-domain (e.g., porn vs. Neutral, violence vs. non-violence), C4Censor presents a unified, balanced, and multi-modal challenge. In benchmarking nine state-of-the-art deep-learning models, even top Vision Transformer variants achieved only 62.1
We present a novel method for tracking the evolution of political dogwhistles—messages which are only understood by a select in-group, while going unnoticed by others (out-group)—in digital environments. Tracking dogwhistles poses a unique empirical challenge due to their reliance on linguistic ambiguity and intentional concealment. To address this, our method combines computational semantics and survey methodology. We model the contextual distribution of dogwhistle terms in online discussion forums, enabling us to infer semantic representations over time, including interpretations not universally recognized by all readers. Diachronic word embeddings are compared with data from a linguistic replacement task that elicits paraphrases reflecting in-group and out-group interpretations. This allows us to track the gradual semantic change of dogwhistles with regard to their in-group and out-group meanings. We demonstrate our method by analyzing the life cycles of four immigration-related dogwhistles across two Swedish online discussion forums over a 23-year period (2000–2022). Our findings reveal different trajectories of semantic change, both across terms and between communities, highlighting the dynamic and context-dependent nature of dogwhistle communication.
In stock markets, prices often respond to specific anchors such as past peaks adjust only gradually to supply–demand imbalances, rather than reflecting fundamentals immediately. These phenomena suggest path dependence, where investor decisions are influenced not only by current market conditions but also by the trajectory of prices leading up to the present. One plausible driver of such path dependence is human behavioral biases, which cause aggregate market outcomes to deviate systematically from rational equilibrium. The challenge in formalizing these biases is that they are inherently context-dependent: their manifestations vary not only with current market conditions but also with factors such as individual trading history or the environment in which the investor operates. To capture the context-dependent nature of behavioral biases and to investigate how such context dependence contributes to the emergence of path-dependent price dynamics in stock markets, we adopt a large language model- (LLM-)augmented agent-based modeling approach. Our analysis proceeds in two stages. Micro analysis: we run controlled trading experiments to test LLMs’ context-dependent behavioral biases. Macro analysis: we introduce the LLM-based agents in an artificial market and conduct multi-agent simulations. These experiments revealed that (1) LLMs’ behavioral biases are context-dependent similar to humans and (2) inclusion of LLMs into artificial market simulations enables the reproduction of path-dependent anomalies in stock prices that conventional agent models had previously failed to capture. Together, these results demonstrate how LLM-based agents can advance constructive modeling of market dynamics.
Possible analogy between traditional scientometrics and altmetrics/scientometrics 2.0 is intriguing. So far, a few evidences are available in regard to the parallelism. More evidence to parallelism might highlight the importance of altmetrics in STI policy/governance, even though its usability as early scholarly impact or societal impact indicator is still debated. This is because altmetrics reflects attention/visibility of research and ensuring early visibility/attention can be crucial for nations and institutions for improving chances for leveraging the transformation potential of their research output, especially the internal research output. For this, the level of dependency of nations and institutions on foreign collaborators for productivity, impact and attention needs to be determined. Recently introduced boost indicators are capable of reflecting the boost in productivity, impact and attention due to foreign collaborations and thereby reflect the level of dependency too. Once level of dependency is known and if boost in attention/altmetrics exhibits parallelism with boost in impact/citations, suitable strategies for enhancing internal scholarly ecosystem of a country can be formulated. In this work, we explore the relationship between indicators related to collaborative boost with respect to citations and altmetrics for (i) 193 UN member countries and (ii) top productive institutions from 4 selected countries (belonging to four categories). At the country-level, a strong positive correlation is found between boost in altmetrics and citations, reinforcing the conjecture about parallelism. At the institutional level, the strength of correlation is found to vary according to the category in which the institution’s country belongs. This highlights the importance of improving necessity of strengthening the scholarly ecosystem of the countries, especially by enhancing institutions’ impact and attention while reducing their dependency on international collaboration. Potential strategies to be adopted by countries belonging to different categories for framing STI Policies and aiding STI governance in this regard are also recommended.
Ranked Choice Voting (RCV) adoption is expanding across U.S. elections, but faces persistent criticism for complexity, strategic manipulation, and ballot exhaustion. We empirically test these concerns on real election data, across three diverse contexts: New York City's 2021 Democratic primaries (54 races), Alaska's 2024 primary-infused statewide elections (52 races), and Portland's 2024 multi-winner City Council elections (4 races). Our algorithmic approach circumvents computational complexity barriers by reducing election instance sizes (via candidate elimination). Our findings reveal that despite its intricate multi-round process and theoretical vulnerabilities, RCV consistently exhibits simple and transparent dynamics in practice, closely mirroring the interpretability of plurality elections. Following RCV adoption, competitiveness increased substantially compared to prior plurality elections, with average margins of victory declining by 9.2 percentage points in NYC and 11.4 points in Alaska. Empirically, complex ballot-addition strategies are not more efficient than simple ones, and ballot exhaustion has minimal impact, altering outcomes in only 3 of 110 elections. These findings demonstrate that RCV delivers measurable democratic benefits while proving robust to ballot-addition manipulation, resilient to ballot exhaustion effects, and maintaining transparent competitive dynamics in practice. The computational framework offers election administrators and researchers tools for immediate election-night analysis and facilitating clearer discourse around election dynamics.
