Decoding Fake Narratives in Spreading Hateful Stories: A Dual-Head RoBERTa Model with Multi-Task Learning

Decoding Fake Narratives in Spreading Hateful Stories: A Dual-Head RoBERTa Model with Multi-Task Learning Yash Bhaskar1 IIIT Hyderabad yash.bhaskar@research.iiit.ac.in Sankalp Bahad1 IIIT Hyderabad sankalp.bahad@research.iiit.ac.in Parameswari Krishnamurthy2 IIIT Hyderabad param.krishna@iiit.ac.in Abstract Social media platforms, while enabling global connectivity, have become hubs for the rapid spread of harmful content, including hate speech and fake narratives (Davidson et al., 2017; Shu et al., 2017). The Faux-Hate shared task focuses on detecting a specific phe- nomenon: the generation of hate speech driven by fake narratives, termed Faux-Hate. Partici- pants are challenged to identify such instances in code-mixed Hindi-English social media text. This paper describes our system developed for the shared task, addressing two primary sub- tasks: (a) Binary Faux-Hate detection, involv- ing fake and hate speech classification, and (b) Target and Severity prediction, categoriz- ing the intended target and severity of hate- ful content. Our approach combines advanced natural language processing techniques with domain-specific pretraining to enhance perfor- mance across both tasks. The system achieved competitive results, demonstrating the efficacy of leveraging multi-task learning for this com- plex problem. 1 Introduction Social media has revolutionized communication, providing unprecedented connectivity across the globe. This increased connectivity, however, has also inadvertently fostered the rapid dissemination of harmful content, including the troubling com- bination of hate speech and fabricated narratives. Hate speech, particularly when intertwined with falsehoods, exacerbates its detrimental impact, fu- eling discrimination, violence, and societal unrest. In response to this growing concern, the Faux- Hate shared task (Biradar et al., 2024a), based on a phenomenon recently characterized and dataset curated by Biradar et al. (Biradar et al., 2024b), in- troduces a unique challenge: identifying and cate- gorizing instances of hate speech generated through fake narratives in code-mixed Hindi-English text. This task emphasizes the importance of detecting and analyzing content that misleads and provokes through a combination of misinformation and hate- ful language. Researchers and practitioners have increasingly turned their attention to understanding and combating these complex phenomena. The shared task comprises two sub-tasks: Task A focuses on binary classification of fake and hate labels, while Task B involves predicting the tar- get and severity of hateful content. This paper describes our system, methodologies, and experi- mental results for both sub-tasks, contributing to the broader effort to address hate speech and fake narratives in multilingual, code-mixed contexts. 2 Related Work The detection of hate speech and misinforma- tion on social media has been a prominent area of research within natural language processing (NLP). Studies have extensively explored tech- niques for identifying hate speech across various languages and platforms (Warner and Hirschberg, 2012), often leveraging machine learning and deep learning approaches. Recent advancements in- clude transformer-based models like BERT (Devlin et al., 2019), RoBERTa, and multilingual BERT (mBERT), which have shown significant success in text classification tasks, including hate speech detection. Fake news and misinformation detection have similarly gained attention (Zubiaga et al., 2018), with methods ranging from linguistic feature anal- ysis to neural network-based classification. The intersection of hate speech and fake narratives, however, remains a relatively unexplored domain, particularly in code-mixed languages like Hindi- English. Prior work in code-mixed text process- ing has highlighted the challenges posed by non- standard grammar, orthographic variations, and the lack of annotated datasets. This shared task builds on these research threads, offering a novel opportunity to investigate Faux- arXiv:2512.16147v1 [cs.CL] 18 Dec 2025 Hate in a multilingual and culturally nuanced con- text. Our approach draws inspiration from prior work in hate speech and fake news detection while tailoring solutions to the unique challenges of the code-mixed Hindi-English dataset provided in this task. 3 Methodology This section outlines the architecture, compo- nents, and training methodology of the dual-head RoBERTa model developed for the Faux-Hate shared task. Our system leverages RoBERTa- base (Liu et al., 2019) as the backbone encoder and extends it with a dual-head classification mechanism for simultaneous hate speech and fake news detection. The architecture adopts a multi- task learning approach (Caruana, 1997), enabling the model to effectively share information across tasks while maintaining task-specific parameteriza- tion through dedicated classification heads. The code for our system, including implementation details and pre-train

This content is AI-processed based on open access ArXiv data.