AMR-GCC: Two-Step Cross-Document Event Factuality Identification on Data Augmentation

This paper focuses on Cross-Document Event Factuality Identification (CEFI) which aims to infer the factual accuracy of a cross-document event from several different document-level texts. As a latest task in Event Factuality Identification (EFI) and an expanded task of Document-level Event Factuality (DEFI), CEFI does not have any corpus or methods, since existing related work is limited to document-level task. Based on these issues, we construct a new corpus, which faces the sparsity of the data and the uneven distribution with different labels limit due to the characteristics of news corpus, and propose Abstract Meaning Representation Graph-based Cross-Document Classification (AMR-GCC) as a two-step method. Combined with additional data generated by fine-tuning GLM-32B in terms of data augmentation, we divide AMR-GCC into two sub tasks to improve the effectiveness of long text processing, i.e., document-level event classification and event collaboration. The results show our method outperforms several SOTAs.