Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yash Bhaskar, Sankalp Bahad, Parameswari Krishnamurthy

TL;DR

The paper tackles detecting hate speech generated by fake narratives in code-mixed Hindi-English text (Faux-Hate) using a dual-head RoBERTa model trained with multi-task learning to perform binary Faux-Hate detection and target/severity classification. It builds on RoBERTa-base with two parallel heads, each featuring progressive dimensionality reduction, layer normalization, GELU activations, and dropout, plus optional residual connections and shared dropout. Experimental results show residual connections improve Task A (F1 up to 0.76) and provide modest gains for Task B, demonstrating the value of multitask sharing in a multilingual, code-mixed setting. The work contributes a practical approach for simultaneous detection and characterization of hateful content embedded in fabricated narratives, with plans for richer data and refined fine-tuning to enhance robustness.

Abstract

Social media platforms, while enabling global connectivity, have become hubs for the rapid spread of harmful content, including hate speech and fake narratives \cite{davidson2017automated, shu2017fake}. The Faux-Hate shared task focuses on detecting a specific phenomenon: the generation of hate speech driven by fake narratives, termed Faux-Hate. Participants 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, involving fake and hate speech classification, and (b) Target and Severity prediction, categorizing the intended target and severity of hateful content. Our approach combines advanced natural language processing techniques with domain-specific pretraining to enhance performance across both tasks. The system achieved competitive results, demonstrating the efficacy of leveraging multi-task learning for this complex problem.

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

TL;DR

The paper tackles detecting hate speech generated by fake narratives in code-mixed Hindi-English text (Faux-Hate) using a dual-head RoBERTa model trained with multi-task learning to perform binary Faux-Hate detection and target/severity classification. It builds on RoBERTa-base with two parallel heads, each featuring progressive dimensionality reduction, layer normalization, GELU activations, and dropout, plus optional residual connections and shared dropout. Experimental results show residual connections improve Task A (F1 up to 0.76) and provide modest gains for Task B, demonstrating the value of multitask sharing in a multilingual, code-mixed setting. The work contributes a practical approach for simultaneous detection and characterization of hateful content embedded in fabricated narratives, with plans for richer data and refined fine-tuning to enhance robustness.

Abstract

Social media platforms, while enabling global connectivity, have become hubs for the rapid spread of harmful content, including hate speech and fake narratives \cite{davidson2017automated, shu2017fake}. The Faux-Hate shared task focuses on detecting a specific phenomenon: the generation of hate speech driven by fake narratives, termed Faux-Hate. Participants 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, involving fake and hate speech classification, and (b) Target and Severity prediction, categorizing the intended target and severity of hateful content. Our approach combines advanced natural language processing techniques with domain-specific pretraining to enhance performance across both tasks. The system achieved competitive results, demonstrating the efficacy of leveraging multi-task learning for this complex problem.

Paper Structure

This paper contains 16 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Base Architecture
  5. Base Encoder
  6. Dual-Head Classification System
  7. Classification Head Architecture
  8. Additional Features
  9. Key Innovations
  10. Experiments and Results
  11. Experimental Setup
  12. Training and Evaluation
  13. Evaluation Metrics
  14. Results
  15. Analysis
  16. ...and 1 more sections

Figures (1)

  • Figure 1: Model architecture