Reasoning over merged semantic graphs across multiple frameworks for fact verification

Abstract

The rapid proliferation of misinformation on the internet has led to an increased need for reliable and efficient fact verification systems. In this thesis, we propose a novel fact verification system that leverages semantic graphs to enhance reasoning in multi-evidence situations. Our system performs reasoning by joining semantic graphs of various retrieved evidence sentences from a trusted fact corpus. Additionally, our system combines semantic graphs from different frameworks to jointly make claim verification predictions, providing a more comprehensive understanding of the underlying facts. We also incorporate a language model-based retrieval model to improve the retrieval of relevant evidence sentences, further enhancing the overall performance. Experimental results on the FEVER dataset demonstrate that our approach outperforms existing fact verification systems, especially in cases where multiple pieces of evidence are required for verification. Ablation studies are conducted to analyze the contributions of different components within our system. This work not only contributes to the advancement of fact verification methods but also lays the groundwork for future research in the area of semantic graph-based reasoning systems.

Date of Award	2023
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Yangqiu SONG (Supervisor)