Fake news detection through feature fusion: Leveraging RoBERTa and knowledge graphs with gating

This dissertation explores feature fusion by combining RoBERTa and Knowledge Graph (KG) techniques using Gated Units to improve the accuracy of fake news detection. In text processing, RoBERTa model is able to understand and classify false content effectively due to its pre-training advantage. On th...

Full description

Saved in:
Bibliographic Details
Main Author: Fang, Zhuohao
Other Authors: Na Jin Cheon
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/181597
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:This dissertation explores feature fusion by combining RoBERTa and Knowledge Graph (KG) techniques using Gated Units to improve the accuracy of fake news detection. In text processing, RoBERTa model is able to understand and classify false content effectively due to its pre-training advantage. On the other hand, knowledge graphs provide rich semantic information in revealing the entities and relationships behind news content, which is crucial for verifying news authenticity. Firstly, this study trains and evaluates the model using the PHEME dataset, and compares the effects of using RoBERTa alone, the knowledge graph alone (processed by TF-IDF and graph neural network (GNN) methods), and their fusion effects. The results show that RoBERTa has the highest accuracy when used alone, emphasizing the power of the pre-trained model for textual semantic understanding. However, when knowledge graph processing is introduced, especially through the GNN approach, it is able to provide a deeper understanding of entity relationships despite longer processing times, which in some cases contributes to improved detection accuracy. Through comparative experiments, this paper confirms the effectiveness of feature fusion strategies in fake news detection and explores the trade-offs between different feature processing methods. In addition, this study points out the limitations of the current approach in terms of data diversity and model generalization, providing directions for improvement in future research. Overall, This dissertation proposes an innovative feature fusion framework that combines features from the text after processing by a pre-trained model (Roberta) and external knowledge from the Knowledge Graph to improve the performance of fake news detection; investigates the Knowledge Graph feature extraction method and improves the utilization of external knowledge by constructing a new graph structure; explores feature fusion strategies and uses the gating unit to improve the ability of model feature fusion. Through these contributions, this research provides new ideas and directions for the field of fake news detection.