Deciphering Rumors: A Multi-Task Learning Approach with Intent-aware Hierarchical Contrastive Learning

Chang Yang, Peng Zhang, Hui Gao, Jing ZhangEMNLP 2024 · DOI

Introduction

Rumor detection on social media is a complex challenge driven by two primary human factors: the diversity of information ("ability to believe") and inherent subjectivity ("willingness to believe"). Users often fall prey to the anchoring effect, evaluating new information through the lens of existing cognitive biases rather than objective facts. Researchers from Tianjin University have introduced IRDNet, a framework designed to move beyond simple text analysis to uncover the latent intent of information disseminators.

The Challenge of Intent in Data Augmentation

Standard machine learning techniques, such as Contrastive Learning (CL), often struggle with rumor detection. Traditional data augmentation methods—like word swapping or sentence cropping—frequently destroy the subtle linguistic cues that signal a user's intent. This leads models to learn incorrect associations. IRDNet addresses this by explicitly modeling the "why" behind a post, ensuring that the semantic "soul" of the message remains intact during training.

Comparison of traditional data augmentation versus ChatGPT-assisted augmentation, highlighting how the latter preserves the original intent.

Methodology: The IRDNet Architecture

IRDNet utilizes a multi-task learning framework that jointly optimizes for rumor classification and latent intent mining. Its architecture is divided into two primary functional modules:

1. Semantic Feature Extraction Module

To handle the unique linguistic style of social media, the model employs BERTweet, a language model pre-trained specifically on Twitter data. This is coupled with BiLSTM and Capsule Networks. While BiLSTM captures sequential patterns, Capsule Networks are utilized to maintain spatial and hierarchical relationships within the text, identifying deep semantic structures that standard convolutional layers might overlook.

2. Intent-aware Hierarchical Contrastive Learning (IHCLM)

This module operates at two levels to ensure robustness:

  • Event-Level CL: High-quality data augmentations are generated using ChatGPT to preserve intent. Additionally, Adversarial Training (via the Fast Gradient Sign Method) is incorporated to protect the model against malicious perturbations.
  • Intent-Level CL: An "intent encoder" utilizes an Intent Matrix representing $K$ latent intentions. This allows the model to cluster tweets with similar subjective motivations while separating those with differing intents, effectively filtering out irrelevant noise.
The complete IRDNet architecture, illustrating the flow from source text to semantic extraction and hierarchical contrastive learning.

Finally, the model uses Classification with Alternating Normalization (CAN) during the prediction stage to refine low-confidence results based on the prior distribution of the training data.

Performance and Key Insights

The authors evaluated IRDNet on the established Twitter15 and Twitter16 benchmarks. The model achieved state-of-the-art accuracy, reaching 91.7% on Twitter15 and 90.9% on Twitter16, outperforming strong baselines like BiGCN and GACL.

Superior Early Detection

A critical finding is IRDNet's effectiveness in the early stages of a rumor's spread. Unlike graph-based models that require extensive propagation trees (comments and shares), IRDNet focuses on the intent within the source tweet. In scenarios with limited training data (a 0.1 training ratio), IRDNet demonstrated an accuracy improvement of 10% to 30% over traditional graph-based neural networks.

The Role of Intent

Ablation studies confirmed that removing the Intent-aware Hierarchical Contrastive Learning module led to a significant accuracy drop (~3%). This underscores that modeling psychological intent—whether to provoke, warn, or mislead—is a vital differentiator in distinguishing factual corrections from malicious fabrications.

Future Directions

While IRDNet sets a new performance ceiling, the researchers acknowledge the computational expense of integrating LLMs and Capsule Networks. Future work will focus on optimizing the model for real-time deployment, expanding to multilingual environments, and adapting to the rapidly evolving nature of online misinformation.

Summary updated

BibTeX
@inproceedings{yang2024deciphering,
  author = {Chang Yang and Peng Zhang and Hui Gao and Jing Zhang},
  title = {{Deciphering Rumors: A Multi-Task Learning Approach with Intent-aware Hierarchical Contrastive Learning}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.emnlp-main.256},
}
Figure