research
List of published and ongoing research work
ongoing
- Discrimination in the Delivery of Betting Advertisements on MetaAatman Vaidya, and Kiran Garimella2026
- Information Seeking Across ChatGPT, Google Search and YouTubeAatman Vaidya, and Kiran Garimella2026
- Modelling the Spread of Toxicity and Exploring its Mitigation on Online Social NetworksAatman Vaidya, Harsh Bhagat, Seema Nagar, and 1 more author2026
Hate speech on online platforms has been credibly linked to multiple instances of real world violence. This calls for an urgent need to understand how toxic content spreads and how it might be mitigated on online social networks, and expectedly has been the topic of extensive research in recent times. Prior work has largely modelled hate through epidemic or spread activation based diffusion models, in which the users are often divided into two categories, hateful or not. In this work, users are treated as transformers of toxicity, based on how they respond to incoming toxicity. Compared with the incoming toxicity, users amplify, attenuate, or replicate (effectively, transform) the toxicity and send it forward. We do a temporal analysis of toxicity on Twitter, Koo and Gab and find that (a) toxicity is not conserved in the network; (b) only a subset of users change behaviour over time; and (c) there is no evidence of homophily among behaviour-changing users. In our model, each user transforms incoming toxicity by applying a "shift" to it prior to sending it forward. Based on this, we develop a network model of toxicity spread that incorporates time-varying behaviour of users. We find that the "shift" applied by a user is dependent on the input toxicity and the category. Based on this finding, we propose an intervention strategy for toxicity reduction. This is simulated by deploying peace-bots. Through experiments on both real-world and synthetic networks, we demonstrate that peace-bot interventions can reduce toxicity, though their effectiveness depends on network structure and placement strategy.
published
- ACL 2026Transfer Learning for Multilingual Polarization DetectionAatman VaidyaIn Workshop on Semantic Evaluation at Association for Computational Linguistics (ACL) 2026
This paper describes our system for Subtask 1 of SemEval-2026 Task 9: POLAR, which focuses on multilingual polarization detection. The task is formulated as a binary classification problem across 22 languages drawnfrom diverse online platforms and real-world events. We investigate three complementary approaches: supervised fine-tuning of multi-lingual encoder-only transformer models, zero-and few-shot classification using large language models (LLMs), and transfer learning from related harmful language tasks such as hate speech, toxicity, abusive language, and gender-based violence. Among the supervised models, mDeBERTa achieved the strongest baseline performance. Prompt-based methods with open-weight LLMs showed limited effectiveness, particularly in zero-shot settings. The best resultswere obtained using transfer learning, where the model was first fine-tuned on related task datasets and then adapted to the polarizationtask, achieving a Macro-F1 score of 0.81. Our findings indicate that supervised multilingualencoders remain highly effective for polarization detection and that incorporating related harmful language tasks can substantially improve performance, especially for nuanced and context-dependent expressions of polarization.
- ICWSM 2026Quantifying the Illicit Ecosystem of Betting Apps in IndiaAatman Vaidya, and Kiran Garimella🏆 Best Paper AwardIn AAAI Conference on Web and Social Media (ICWSM) 2026
Online betting and gambling apps in India have expanded rapidly, alongside growing concern about financial loss, debt stress, and addictive use patterns. Yet the ecosystem is difficult to quantify because recruitment and harm are observed in different places: users are often acquired through social media promotion, while harms become visible later inside apps and in user complaints. We address this measurement gap with a mixed-method, multi-source study that links promotion to downstream experience. We compile three complementary datasets. First, we collect and analyze tens of thousands of betting-related advertisements from Meta’s Ad Library using an extensive keyword strategy to measure scale and characterize persuasive frames. Second, we gather a purposive sample of organic Instagram posts from ten betting-linked hashtags to study how similar narratives circulate outside formal advertising, including through surrogate sports pages and influencer-style content. Third, we analyze over 300,000 Google Play reviews for a set of betting apps, using topic modeling to extract recurring user-reported problems that reflect a harm surface including financial loss, withdrawal friction, and customer support failure. We connect these layers by constructing a shared narrative codebook for paid and organic promotion and mapping those recruitment narratives to review topics. Across sources, we find a consistent mismatch between what is promised at recruitment and what users report after adoption. Paid ads frequently frame betting as simple, quick, and highly winnable, while reviews repeatedly describe difficulty winning, blocked or delayed withdrawals, unclear rules, and perceived extractive design. Organic promotion often uses more coded and informal presentation than official ads, potentially reducing detectability while funneling users toward the same apps and referral pathways. Together, these results provide one of the first large-scale, cross-source measurements of India’s online betting promotion ecosystem and its associated user-reported harms, and they offer a general approach for studying how potentially harmful digital services sustain growth through mainstream platforms even under evolving regulation.
- ArXivAnalysis of Indic Language Capabilities in LLMsIn 2025
This report evaluates the performance of text-in text-out Large Language Models (LLMs) to understand and generate Indic languages. This evaluation is used to identify and prioritize Indic languages suited for inclusion in safety benchmarks. We conduct this study by reviewing existing evaluation studies and datasets; and a set of twenty-eight LLMs that support Indic languages. We analyze the LLMs on the basis of the training data, license for model and data, type of access and model developers. We also compare Indic language performance across evaluation datasets and find that significant performance disparities in performance across Indic languages. Hindi is the most widely represented language in models. While model performance roughly correlates with number of speakers for the top five languages, the assessment after that varies.
- NAACL 2024The Uli Dataset: An Exercise in Experience Led Annotation of oGBVArnav Arora, Maha Jinadoss, Cheshta Arora, and 5 more authors🏆 Outstanding Paper AwardIn Workshop on Online Abuse and Harms at Annual Conference of the North American Chapter of the Association for Computational Linguistics 2024
Online gender-based violence has grown concomitantly with the adoption of the internet and social media. Its effects are worse in the Global majority where many users use social media in languages other than English. The scale and volume of conversations on the internet have necessitated the need for automated detection of hate speech and, more specifically, gendered abuse. There is, however, a lack of language-specific and contextual data to build such automated tools. In this paper, we present a dataset on gendered abuse in three languages- Hindi, Tamil and Indian English. The dataset comprises of tweets annotated along three questions pertaining to the experience of gender abuse, by experts who identify as women or a member of the LGBTQIA+ community in South Asia. Through this dataset, we demonstrate a participatory approach to creating datasets that drive AI systems.
- CODS-COMADAnalysing the Spread of Toxicity on TwitterAatman Vaidya, Seema Nagar, and Amit A NanavatiIn Proceedings of the 7th Joint International Conference on Data Science & Management of Data (11th ACM IKDD CODS and 29th COMAD) 2024
The spread of hate speech on social media platforms has become a rising concern in recent years. Understanding the spread of hate is crucial for mitigating its harmful effects and fostering a healthier online environment. In this paper, we propose a new model to capture the evolution of toxicity in a network – if a tweet with a certain toxicity (hatefulness) is posted, how much toxic a social network will become after a given number of rounds. We compute a toxicity score for each tweet, indicating the extent of the hatefulness of that tweet. Toxicity spread has not been adequately addressed in the existing literature. The two popular paradigms for modelling information spread, namely the Susceptible-Infected-Recovered (SIR) and its variants, as well as the spreading-activation models (SPA), are not suitable for modelling toxicity spread. The first paradigm employs a threshold and categorizes tweets as either toxic or non-toxic, while the second paradigm treats hate as energy and applies energy-conversion principles to model its propagation. Through analysis of a Twitter dataset consisting of 19.58 million tweets, we observe that the total toxicity, as well as the average toxicity of original tweets and retweets in the network, does not remain constant but rather increases over time. In this paper, we propose a new method for toxicity spread. First, we categorize users into three distinct groups: Amplifiers, Attenuators, and Copycats. These categories are assigned based on the exchange of toxicity by a user, with Amplifiers sending out more toxicity than they receive, Attenuators experiencing a higher influx of toxicity compared to what they generate, and Copycats simply mirroring the hate they receive. We perform extensive experimentation on Barabási–Albert (BA) graphs, as well as subgraphs extracted from the Twitter dataset. Our model is able to replicate the patterns of toxicity.