Enhancing rumour classification with target-based dual emotion using LLM

• Main Author: Li, Yuanhang
• Other Authors: Na Jin Cheon
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Social Sciences; Rumour classification; Target-based dual emotion; LLM
• Online Access: https://hdl.handle.net/10356/181072
• Institution: Nanyang Technological University

Emotion plays a significant role in detecting online rumours. While existing methods focus on publisher emotion, the emotions of readers (i.e., reader emo tion) are often overlooked. Moreover, the relationship between publisher emotion and reader emotion (i.e., dual emotion) and its manifestation in rumours re mains underexplored. This paper introduces a refined target-based dual emotion approach, validated using the RumourEval-19 and PHEME datasets. A K-S test confirms that different rumour types exhibit distinct distributions of target-based dual emotion. We integrate this feature with the large language model FLAN T5 to evaluate performance in rumour detection and credibility assessment. Our results show that target-based category dual emotion achieves a macro F1 score of 0.420 in credibility assessment (classify the rumour is true, false or unveri- fied ), while target-based specific dual emotion reaches 88.0% accuracy and an F1 score of 0.863 in rumour detection. Additionally, we demonstrate that the model with specific target-based emotion features performs well in real-world rumour outbreak scenarios. Overall, our target-based dual emotion input allows the large language model to achieve results close to, and sometimes surpass ing, state-of-the-art models. This highlights the potential of target-based dual emotion to enhance performance in rumour detection and credibility assessment tasks, providing a robust solution for combating misinformation.