Content moderation systems are crucial in Online Social Networks (OSNs). Indeed, their role is to keep platforms and their users safe from malicious activities. However, there is an emerging consensus that such systems are unfair to fragile users and minorities. Furthermore, content moderation systems are difficult to personalize and lack effective communication between users and platforms. In this context, we propose an enhancement of the current framework of content moderation, integrating Large Language Models (LLMs) in the enforcing pipeline.
Many countries, including US and Brazil, have laws to guarantee children’s rights against exposure, directly and indirectly, to advertisements that could harm their education. However, given the facilities of creating an online ad on social media and the high volume of posts that could carry child advertising, monitoring compliance with those laws has not been trivial to be done manually. Some influencers use online social media platforms, particularly Instagram, to publish advertisements for children, their parents, and relatives. In this work, we propose a machine learning-based classifier to indicate child advertising on social media using a set of posts manually labeled by specialists with content published by known influencers on Instagram. Moreover, we characterized child advertising in 2019, 2020, and 2021 through the lens of mentions (@) and hashtag (#) usage on Instagram posts. We expect that researchers, governments, non-government organizations, and the general public could use help to improve this classifier and the proposed methodology to build a ground truth to train it to enhance the identification of online child advertising. In contrast, our characterization sheds light on influencer and brand behavior on Instagram. Among other preliminary results, we show that the two best classifiers to detect child advertising on Instagram were Naive Bayes and CNN.
Offering a viable alternative architecture to centrally-controlled global digital platforms for social networking is an open challenge. Here we present a grassroots architecture for serverless, permissionless, peer-to-peer social networks termed grassroots social networking. The architecture is geared for roaming (address-changing) agents communicating over an unreliable network, e.g., smartphones communicating via UDP. The architecture incorporates (i) a decentralized social graph, where each member controls, maintains and stores only their local neighbourhood in the graph; (ii) member-created feeds, with authors and followers; and (iii) a grassroots dissemination protocol, in which communication occurs only along the edges of the social graph. The architecture realizes these components using the blocklace data structure – a distributed partially-ordered counterpart of the replicated totally-ordered blockchain. We provide two example grassroots social networking protocols—Twitter/LinkedIn-like and WhatsApp-like—and address their security (safety, liveness and privacy), spam/deep-fake resistance, and implementation, demonstrating how centrally-controlled social networks could be supplanted by a grassroots architecture.
This paper presents a novel dataset of 200k YouTube comments from 468 videos across 109 channels in four content categories: Entertainment, Gaming, People & Blogs, and Science & Technology. We applied state-of-the-art NLP methods to augment the dataset with sexism-related features such as sentiment, toxicity, offensiveness, and hate speech. These features can assist manual content analyses and enable automated analysis of sexism in online platforms. Furthermore, we develop an annotation framework inspired by the Ambivalent Sexism Theory to promote a nuanced understanding of how comments relate to the gender of content creators. We release a small sample of comments annotated using this framework. Our dataset analysis confirms that female content creators receive more sexist and hateful comments than their male counterparts, underscoring the need for further research and intervention in addressing online sexism.
Web3 social media strives to eliminate the need for centralized management by building upon on technologies such public ledgers and smart contracts. This has generated significant hype, but creates many notable challenges, many of which remain unaddressed. One such challenge is content moderation. Specifically, the immutable nature of blockchain means that it becomes impossible to retrospectively delete posts. This means that illegal content posts cannot be moderated or removed. In an attempt to overcome this, Web3 platforms, such as memo.cash, allow users to filter out posts from their personal timelines. Taking memo.cash as a use case, the goal of this work is to study the efficacy of the approach, and identify associated challenges. A particularly unique feature of memo.cash is that users must pay money (satoshi) for each social action (e.g., posting and blocking other users). We conjecture that this may impact the nature of moderation, particularly among poorer users. To explore this, we gather data from memo.cash covering 24K users, 317K posts, and 2.57M user actions. We investigate how the need to pay may impact the moderation system and propose potential solutions to address the challenges that arise.
Fake news is a major challenge in social media, particularly in the health domain where it can lead to severe consequences for both individuals and society as a whole. To contribute to combating this problem, we present a novel solution for improving the accuracy of detecting fake health news, utilizing a fine-tuned BERT model that integrates both user- and content-related socio-contextual information. Specifically, this information is combined with the textual content itself to form a socio-contextual input sequence for the BERT model. By fine-tuning such a model with respect to the health misinformation detection task, the resulting classifier can accurately predict the category to which each piece of content belongs, i.e., either “real health news” or “fake health news”. We validate our solution through a series of experiments conducted on distinct publicly available datasets constituted by health-related tweets. These results illustrate the superiority of the proposed solution compared to the standard BERT baseline model and other advanced models. Indeed, they show that the integration of socio-contextual information in the detection process positively contributes to increasing the overall accuracy of the fake health news detection task. The study also suggests, in a preliminary way, how such information could be used for the explainability of the model itself.