Nishanth Sastry

Wenjie Yin, Vibhor Agarwal, Aiqi Jiang et al.

Supervised approaches generally rely on majority-based labels. However, it is hard to achieve high agreement among annotators in subjective tasks such as hate speech detection. Existing neural network models principally regard labels as categorical variables, while ignoring the semantic information in diverse label texts. In this paper, we propose AnnoBERT, a first-of-its-kind architecture integrating annotator characteristics and label text with a transformer-based model to detect hate speech, with unique representations based on each annotator's characteristics via Collaborative Topic Regression (CTR) and integrate label text to enrich textual representations. During training, the model associates annotators with their label choices given a piece of text; during evaluation, when label information is not available, the model predicts the aggregated label given by the participating annotators by utilising the learnt association. The proposed approach displayed an advantage in detecting hate speech, especially in the minority class and edge cases with annotator disagreement. Improvement in the overall performance is the largest when the dataset is more label-imbalanced, suggesting its practical value in identifying real-world hate speech, as the volume of hate speech in-the-wild is extremely small on social media, when compared with normal (non-hate) speech. Through ablation studies, we show the relative contributions of annotator embeddings and label text to the model performance, and tested a range of alternative annotator embeddings and label text combinations.

Vibhor Agarwal, Yu Chen, Nishanth Sastry

Hate speech has become pervasive in today's digital age. Although there has been considerable research to detect hate speech or generate counter speech to combat hateful views, these approaches still cannot completely eliminate the potential harmful societal consequences of hate speech -- hate speech, even when detected, can often not be taken down or is often not taken down enough; and hate speech unfortunately spreads quickly, often much faster than any generated counter speech. This paper investigates a relatively new yet simple and effective approach of suggesting a rephrasing of potential hate speech content even before the post is made. We show that Large Language Models (LLMs) perform well on this task, outperforming state-of-the-art baselines such as BART-Detox. We develop 4 different prompts based on task description, hate definition, few-shot demonstrations and chain-of-thoughts for comprehensive experiments and conduct experiments on open-source LLMs such as LLaMA-1, LLaMA-2 chat, Vicuna as well as OpenAI's GPT-3.5. We propose various evaluation metrics to measure the efficacy of the generated text and ensure the generated text has reduced hate intensity without drastically changing the semantic meaning of the original text. We find that LLMs with a few-shot demonstrations prompt work the best in generating acceptable hate-rephrased text with semantic meaning similar to the original text. Overall, we find that GPT-3.5 outperforms the baseline and open-source models for all the different kinds of prompts. We also perform human evaluations and interestingly, find that the rephrasings generated by GPT-3.5 outperform even the human-generated ground-truth rephrasings in the dataset. We also conduct detailed ablation studies to investigate why LLMs work satisfactorily on this task and conduct a failure analysis to understand the gaps.

Vibhor Agarwal, Yiqiao Jin, Mohit Chandra et al. · gatech

Large language models (LLMs) are starting to complement traditional information seeking mechanisms such as web search. LLM-powered chatbots like ChatGPT are gaining prominence among the general public. AI chatbots are also increasingly producing content on social media platforms. However, LLMs are also prone to hallucinations, generating plausible yet factually incorrect or fabricated information. This becomes a critical problem when laypeople start seeking information about sensitive issues such as healthcare. Existing works in LLM hallucinations in the medical domain mainly focus on testing the medical knowledge of LLMs through standardized medical exam questions which are often well-defined and clear-cut with definitive answers. However, these approaches may not fully capture how these LLMs perform during real-world interactions with patients. This work conducts a pioneering study on hallucinations in LLM-generated responses to real-world healthcare queries from patients.We introduce MedHalu, a novel medical hallucination benchmark featuring diverse health-related topics and hallucinated responses from LLMs, with detailed annotation of the hallucination types and text spans. We also propose MedHaluDetect, a comprehensive framework for evaluating LLMs' abilities to detect hallucinations. Furthermore, we study the vulnerability to medical hallucinations among three groups -- medical experts, LLMs, and laypeople. Notably, LLMs significantly underperform human experts and, in some cases, even laypeople in detecting medical hallucinations. To improve hallucination detection, we propose an expert-in-the-loop approach that integrates expert reasoning into LLM inputs, significantly improving hallucination detection for all LLMs, including a 6.3% macro-F1 improvement for GPT-4. Our code and dataset are available at https://netsys.surrey.ac.uk/datasets/medhalu/.

Vibhor Agarwal, Anthony P. Young, Sagar Joglekar et al.

Online forums that allow for participatory engagement between users have been transformative for the public discussion of many important issues. However, such conversations can sometimes escalate into full-blown exchanges of hate and misinformation. Existing approaches in natural language processing (NLP), such as deep learning models for classification tasks, use as inputs only a single comment or a pair of comments depending upon whether the task concerns the inference of properties of the individual comments or the replies between pairs of comments, respectively. But in online conversations, comments and replies may be based on external context beyond the immediately relevant information that is input to the model. Therefore, being aware of the conversations' surrounding contexts should improve the model's performance for the inference task at hand. We propose GraphNLI, a novel graph-based deep learning architecture that uses graph walks to incorporate the wider context of a conversation in a principled manner. Specifically, a graph walk starts from a given comment and samples "nearby" comments in the same or parallel conversation threads, which results in additional embeddings that are aggregated together with the initial comment's embedding. We then use these enriched embeddings for downstream NLP prediction tasks that are important for online conversations. We evaluate GraphNLI on two such tasks - polarity prediction and misogynistic hate speech detection - and found that our model consistently outperforms all relevant baselines for both tasks. Specifically, GraphNLI with a biased root-seeking random walk performs with a macro-F1 score of 3 and 6 percentage points better than the best-performing BERT-based baselines for the polarity prediction and hate speech detection tasks, respectively.

Vahid Ghafouri, Vibhor Agarwal, Yong Zhang et al.

The introduction of ChatGPT and the subsequent improvement of Large Language Models (LLMs) have prompted more and more individuals to turn to the use of ChatBots, both for information and assistance with decision-making. However, the information the user is after is often not formulated by these ChatBots objectively enough to be provided with a definite, globally accepted answer. Controversial topics, such as "religion", "gender identity", "freedom of speech", and "equality", among others, can be a source of conflict as partisan or biased answers can reinforce preconceived notions or promote disinformation. By exposing ChatGPT to such debatable questions, we aim to understand its level of awareness and if existing models are subject to socio-political and/or economic biases. We also aim to explore how AI-generated answers compare to human ones. For exploring this, we use a dataset of a social media platform created for the purpose of debating human-generated claims on polemic subjects among users, dubbed Kialo. Our results show that while previous versions of ChatGPT have had important issues with controversial topics, more recent versions of ChatGPT (gpt-3.5-turbo) are no longer manifesting significant explicit biases in several knowledge areas. In particular, it is well-moderated regarding economic aspects. However, it still maintains degrees of implicit libertarian leaning toward right-winged ideals which suggest the need for increased moderation from the socio-political point of view. In terms of domain knowledge on controversial topics, with the exception of the "Philosophical" category, ChatGPT is performing well in keeping up with the collective human level of knowledge. Finally, we see that sources of Bing AI have slightly more tendency to the center when compared to human answers. All the analyses we make are generalizable to other types of biases and domains.

Pushkal Agarwal, Nishanth Sastry, Edward Wood

In the second half of the 20th century, Parliament allowed broadcasters to transmit radio and eventually television coverage of debates and meetings of select committees. More recently, in an effort to further improve transparency and citizen engagement, the UK Parliament started publishing videos of these debates and meetings itself, and tweeting details of debates as they happened. In this paper, we attempt to characterise how people engage with video data of Parliamentary debates by using more than two years of Google Analytics data around these videos. We analyse the patterns of engagement - how do they land on a particular video? How do they hear about this video, i.e., what is the (HTTP) referrer website that led to the user clicking on the video? Once a user lands on a video, how do they engage with it? For how long is the video played? What is the next destination? etc. Answering these questions is an important first step towards understanding why and how people use Parliamentary videos, and therefore, how the video delivery platform should be adapted and personalised for the needs of the citizens of the country. Taking inspiration from An, Kwak, and Jansen (2017), we employ Non-Negative Matrix Factorization (NMF) (Lee and Seung, 1999) on the video views matrix to identify different archetypes of users, and identify archetypes. A deeper examination of the archetypes we find reveals that they are primarily distinguished by how they land on the video page: Search (i.e., through a search engine), Referral (i.e., from other Parliamentary websites), Direct (i.e., through a direct link, which is embedded on another website), Social (i.e., through a social platform such as Facebook or Twitter) and Others.

Vibhor Agarwal, Yu Chen, Nishanth Sastry

Online conversation understanding is an important yet challenging NLP problem which has many useful applications (e.g., hate speech detection). However, online conversations typically unfold over a series of posts and replies to those posts, forming a tree structure within which individual posts may refer to semantic context from higher up the tree. Such semantic cross-referencing makes it difficult to understand a single post by itself; yet considering the entire conversation tree is not only difficult to scale but can also be misleading as a single conversation may have several distinct threads or points, not all of which are relevant to the post being considered. In this paper, we propose a Graph-based Attentive Semantic COntext Modeling (GASCOM) framework for online conversation understanding. Specifically, we design two novel algorithms that utilise both the graph structure of the online conversation as well as the semantic information from individual posts for retrieving relevant context nodes from the whole conversation. We further design a token-level multi-head graph attention mechanism to pay different attentions to different tokens from different selected context utterances for fine-grained conversation context modeling. Using this semantic conversational context, we re-examine two well-studied problems: polarity prediction and hate speech detection. Our proposed framework significantly outperforms state-of-the-art methods on both tasks, improving macro-F1 scores by 4.5% for polarity prediction and by 5% for hate speech detection. The GASCOM context weights also enhance interpretability.

Shalini Jangra, Suparna De, Nishanth Sastry et al.

Social platforms such as Reddit have a network of communities of shared interests, with a prevalence of posts and comments from which one can infer users' Personal Information Identifiers (PIIs). While such self-disclosures can lead to rewarding social interactions, they pose privacy risks and the threat of online harms. Research into the identification and retrieval of such risky self-disclosures of PIIs is hampered by the lack of open-source labeled datasets. To foster reproducible research into PII-revealing text detection, we develop a novel methodology to create synthetic equivalents of PII-revealing data that can be safely shared. Our contributions include creating a taxonomy of 19 PII-revealing categories for vulnerable populations and the creation and release of a synthetic PII-labeled multi-text span dataset generated from 3 text generation Large Language Models (LLMs), Llama2-7B, Llama3-8B, and zephyr-7b-beta, with sequential instruction prompting to resemble the original Reddit posts. The utility of our methodology to generate this synthetic dataset is evaluated with three metrics: First, we require reproducibility equivalence, i.e., results from training a model on the synthetic data should be comparable to those obtained by training the same models on the original posts. Second, we require that the synthetic data be unlinkable to the original users, through common mechanisms such as Google Search. Third, we wish to ensure that the synthetic data be indistinguishable from the original, i.e., trained humans should not be able to tell them apart. We release our dataset and code at https://netsys.surrey.ac.uk/datasets/synthetic-self-disclosure/ to foster reproducible research into PII privacy risks in online social media.

Aravindh Raman, Sagar Joglekar, Emiliano De Cristofaro et al.

The Decentralised Web (DW) has recently seen a renewed momentum, with a number of DW platforms like Mastodon, Peer-Tube, and Hubzilla gaining increasing traction. These offer alternatives to traditional social networks like Twitter, YouTube, and Facebook, by enabling the operation of web infrastructure and services without centralised ownership or control. Although their services differ greatly, modern DW platforms mostly rely on two key innovations: first, their open source software allows anybody to setup independent servers ("instances") that people can sign-up to and use within a local community; and second, they build on top of federation protocols so that instances can mesh together, in a peer-to-peer fashion, to offer a globally integrated platform. In this paper, we present a measurement-driven exploration of these two innovations, using a popular DW microblogging platform (Mastodon) as a case study. We focus on identifying key challenges that might disrupt continuing efforts to decentralise the web, and empirically highlight a number of properties that are creating natural pressures towards recentralisation. Finally, our measurements shed light on the behaviour of both administrators (i.e., people setting up instances) and regular users who sign-up to the platforms, also discussing a few techniques that may address some of the issues observed.

Zeqiang Wang, Xinyue Wu, Chenxi Li et al.

Large language models and text encoders increasingly power web-based information systems in the social sciences, including digital libraries, data catalogues, and search interfaces used by researchers, policymakers, and civil society. Full fine-tuning is often computationally and energy intensive, which can be prohibitive for smaller institutions and non-profit organizations in the Web4Good ecosystem. Parameter-Efficient Fine-Tuning (PEFT), especially Low-Rank Adaptation (LoRA), reduces this cost by updating only a small number of parameters. We show that the standard LoRA update $ΔW = BA^\top$ has geometric drawbacks: gauge freedom, scale ambiguity, and a tendency toward rank collapse. We introduce OrthoGeoLoRA, which enforces an SVD-like form $ΔW = BΣA^\top$ by constraining the low-rank factors to be orthogonal (Stiefel manifold). A geometric reparameterization implements this constraint while remaining compatible with standard optimizers such as Adam and existing fine-tuning pipelines. We also propose a benchmark for hierarchical concept retrieval over the European Language Social Science Thesaurus (ELSST), widely used to organize social science resources in digital repositories. Experiments with a multilingual sentence encoder show that OrthoGeoLoRA outperforms standard LoRA and several strong PEFT variants on ranking metrics under the same low-rank budget, offering a more compute- and parameter-efficient path to adapt foundation models in resource-constrained settings.

Prince Bhardwaj, Nishanth Sastry

Passkeys -- discoverable WebAuthn credentials synchronised across devices are widely promoted as the future of passwordless authentication. Built on the FIDO2 standard, they eliminate shared secrets and resist phishing while offering usability through platform credential managers. Since their introduction in 2022, major vendors have integrated passkeys into operating systems and browsers, and prominent websites have announced support. Yet the true extent of adoption across the broader web remains unknown. Measuring this is challenging because websites implement passkeys in heterogeneous ways. Some expose explicit "Sign in with passkey" buttons, others hide options under multi-step flows or rely on conditional mediation, and many adopt external mechanisms such as JavaScript libraries or OAuth-based identity providers. There is no standardised discovery endpoint, and dynamic, JavaScript-heavy pages complicate automated detection. This paper makes two contributions. First, we present Fidentikit, a browser-based crawler implementing 43 heuristics across five categories - UI elements, DOM structures, WebAuthn API calls, network patterns, and library detection - developed through iterative refinement over 1,500 sites. Second, we apply Fidentikit to the top 100,000 Tranco-ranked domains, producing the first large-scale census of passkey adoption. Our results show adoption strongly correlates with site popularity and often depends on external identity providers rather than native implementations.

Vibhor Agarwal, Aravindh Raman, Nishanth Sastry et al.

The recent development of decentralised and interoperable social networks (such as the "fediverse") creates new challenges for content moderators. This is because millions of posts generated on one server can easily "spread" to another, even if the recipient server has very different moderation policies. An obvious solution would be to leverage moderation tools to automatically tag (and filter) posts that contravene moderation policies, e.g. related to toxic speech. Recent work has exploited the conversational context of a post to improve this automatic tagging, e.g. using the replies to a post to help classify if it contains toxic speech. This has shown particular potential in environments with large training sets that contain complete conversations. This, however, creates challenges in a decentralised context, as a single conversation may be fragmented across multiple servers. Thus, each server only has a partial view of an entire conversation because conversations are often federated across servers in a non-synchronized fashion. To address this, we propose a decentralised conversation-aware content moderation approach suitable for the fediverse. Our approach employs a graph deep learning model (GraphNLI) trained locally on each server. The model exploits local data to train a model that combines post and conversational information captured through random walks to detect toxicity. We evaluate our approach with data from Pleroma, a major decentralised and interoperable micro-blogging network containing 2 million conversations. Our model effectively detects toxicity on larger instances, exclusively trained using their local post information (0.8837 macro-F1). Our approach has considerable scope to improve moderation in decentralised and interoperable social networks such as Pleroma or Mastodon.

Suparna De, Ionut Bostan, Nishanth Sastry

Recent studies have outlined the accessibility challenges faced by blind or visually impaired, and less-literate people, in interacting with social networks, in-spite of facilitating technologies such as monotone text-to-speech (TTS) screen readers and audio narration of visual elements such as emojis. Emotional speech generation traditionally relies on human input of the expected emotion together with the text to synthesise, with additional challenges around data simplification (causing information loss) and duration inaccuracy, leading to lack of expressive emotional rendering. In real-life communications, the duration of phonemes can vary since the same sentence might be spoken in a variety of ways depending on the speakers' emotional states or accents (referred to as the one-to-many problem of text to speech generation). As a result, an advanced voice synthesis system is required to account for this unpredictability. We propose an end-to-end context-aware Text-to-Speech (TTS) synthesis system that derives the conveyed emotion from text input and synthesises audio that focuses on emotions and speaker features for natural and expressive speech, integrating advanced natural language processing (NLP) and speech synthesis techniques for real-time applications. Our system also showcases competitive inference time performance when benchmarked against the state-of-the-art TTS models, making it suitable for real-time accessibility applications.

Vibhor Agarwal, Sagar Joglekar, Anthony P. Young et al.

Online forums that allow participatory engagement between users have been transformative for public discussion of important issues. However, debates on such forums can sometimes escalate into full blown exchanges of hate or misinformation. An important tool in understanding and tackling such problems is to be able to infer the argumentative relation of whether a reply is supporting or attacking the post it is replying to. This so called polarity prediction task is difficult because replies may be based on external context beyond a post and the reply whose polarity is being predicted. We propose GraphNLI, a novel graph-based deep learning architecture that uses graph walk techniques to capture the wider context of a discussion thread in a principled fashion. Specifically, we propose methods to perform root-seeking graph walks that start from a post and captures its surrounding context to generate additional embeddings for the post. We then use these embeddings to predict the polarity relation between a reply and the post it is replying to. We evaluate the performance of our models on a curated debate dataset from Kialo, an online debating platform. Our model outperforms relevant baselines, including S-BERT, with an overall accuracy of 83%.

Pushkal Agarwal, Aravindh Raman, Damilola Ibosiola et al.

WhatsApp is a popular messaging app used by over a billion users around the globe. Due to this popularity, understanding misbehavior on WhatsApp is an important issue. The sending of unwanted junk messages by unknown contacts via WhatsApp remains understudied by researchers, in part because of the end-to-end encryption offered by the platform. We address this gap by studying junk messaging on a multilingual dataset of 2.6M messages sent to 5K public WhatsApp groups in India. We characterise both junk content and senders. We find that nearly 1 in 10 messages is unwanted content sent by junk senders, and a number of unique strategies are employed to reflect challenges faced on WhatsApp, e.g., the need to change phone numbers regularly. We finally experiment with on-device classification to automate the detection of junk, whilst respecting end-to-end encryption.

Pushkal Agarwal, Kiran Garimella, Sagar Joglekar et al.

Social media has been on the vanguard of political information diffusion in the 21st century. Most studies that look into disinformation, political influence and fake-news focus on mainstream social media platforms. This has inevitably made English an important factor in our current understanding of political activity on social media. As a result, there has only been a limited number of studies into a large portion of the world, including the largest, multilingual and multi-cultural democracy: India. In this paper we present our characterisation of a multilingual social network in India called ShareChat. We collect an exhaustive dataset across 72 weeks before and during the Indian general elections of 2019, across 14 languages. We investigate the cross lingual dynamics by clustering visually similar images together, and exploring how they move across language barriers. We find that Telugu, Malayalam, Tamil and Kannada languages tend to be dominant in soliciting political images (often referred to as memes), and posts from Hindi have the largest cross-lingual diffusion across ShareChat (as well as images containing text in English). In the case of images containing text that cross language barriers, we see that language translation is used to widen the accessibility. That said, we find cases where the same image is associated with very different text (and therefore meanings). This initial characterisation paves the way for more advanced pipelines to understand the dynamics of fake and political content in a multi-lingual and non-textual setting.

Pushkal Agarwal, Sagar Joglekar, Panagiotis Papadopoulos et al.

Websites with hyper-partisan, left or right-leaning focus offer content that is typically biased towards the expectations of their target audience. Such content often polarizes users, who are repeatedly primed to specific (extreme) content, usually reflecting hard party lines on political and socio-economic topics. Though this polarization has been extensively studied with respect to content, it is still unknown how it associates with the online tracking experienced by browsing users, especially when they exhibit certain demographic characteristics. For example, it is unclear how such websites enable the ad-ecosystem to track users based on their gender or age. In this paper, we take a first step to shed light and measure such potential differences in tracking imposed on users when visiting specific party-line's websites. For this, we design and deploy a methodology to systematically probe such websites and measure differences in user tracking. This methodology allows us to create user personas with specific attributes like gender and age and automate their browsing behavior in a consistent and repeatable manner. Thus, we systematically study how personas are being tracked by these websites and their third parties, especially if they exhibit particular demographic properties. Overall, we test 9 personas on 556 hyper-partisan websites and find that right-leaning websites tend to track users more intensely than left-leaning, depending on user demographics, using both cookies and cookie synchronization methods and leading to more costly delivered ads.

Xuehui Hu, Nishanth Sastry

The recently introduced General Data Protection Regulation (GDPR) requires that when obtaining information online that could be used to identify individuals, their consents must be obtained. Among other things, this affects many common forms of cookies, and users in the EU have been presented with notices asking their approvals for data collection. This paper examines the prevalence of third party cookies before and after GDPR by using two datasets: accesses to top 500 websites according to Alexa.com, and weekly data of cookies placed in users' browsers by websites accessed by 16 UK and China users across one year. We find that on average the number of third parties dropped by more than 10% after GDPR, but when we examine real users' browsing histories over a year, we find that there is no material reduction in long-term numbers of third party cookies, suggesting that users are not making use of the choices offered by GDPR for increased privacy. Also, among websites which offer users a choice in whether and how they are tracked, accepting the default choices typically ends up storing more cookies on average than on websites which provide a notice of cookies stored but without giving users a choice of which cookies, or those that do not provide a cookie notice at all. We also find that top non-EU websites have fewer cookie notices, suggesting higher levels of tracking when visiting international sites. Our findings have deep implications both for understanding compliance with GDPR as well as understanding the evolution of tracking on the web.

Changtao Zhong, Hau-wen Chan, Dmytro Karamshuk et al.

This paper investigates when users create profiles in different social networks, whether they are redundant expressions of the same persona, or they are adapted to each platform. Using the personal webpages of 116,998 users on About.me, we identify and extract matched user profiles on several major social networks including Facebook, Twitter, LinkedIn, and Instagram. We find evidence for distinct site-specific norms, such as differences in the language used in the text of the profile self-description, and the kind of picture used as profile image. By learning a model that robustly identifies the platform given a user's profile image (0.657--0.829 AUC) or self-description (0.608--0.847 AUC), we confirm that users do adapt their behaviour to individual platforms in an identifiable and learnable manner. However, different genders and age groups adapt their behaviour differently from each other, and these differences are, in general, consistent across different platforms. We show that differences in social profile construction correspond to differences in how formal or informal the platform is.