A multi-view heterogeneous and extractive graph attention network for evidential document-level event factuality identification

Document-level Event Factuality Identification (DEFI) has always been the focus of Information Extraction and Information Credibility. This paper studies a novel task named Evidential Document-level Event Factuality Identification (EvDEFI), which aims to predict the factual nature of an event and extract evidential sentences from the document precisely. However, the existing methods usually rely heavily on specific annotated texts, lack interpretability, or utilize coarse-grained solutions.

To solve the problems, a research team led by Zhong QIAN published their new research on 15 June 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a novel fine-grained and interpretable model named Heterogeneous and Extractive Graph Attention neTwork (HEGAT). This model is verified and tested on a refined corpus called EB-DEF-v2. Compared with the existing baselines, the proposed method can achieve SOTA and validate the interpretability of the task.

In the research, they analyze the main challenges and problems that lie in the extraction of sentence-level event mentions and corresponding speculation/negation semantics. Therefore, the proposed HEGAT model considers multi-granularity and multi-view fusion involving token-level, sentence-level and document-level on EvDEFI task. In order to enhance the interpretability, HEGAT also integrates lexical features including part-of-speeches, named entities, speculative and negative cues, and designing a span extraction mechanism. By multi-view graph attentions, the model can update the states of events and sentences based on tokens and lexical features from both local and global levels. Finally, they expand the original corpus into a new one called EB-DEF-v2, on which HEGAT can outperform several state-of-the-art baselines, verifying the effectiveness of the proposed model.

As for the future work, they will explore more challenging versions of EvDEFI, including the situations of multi-event cross-document, or multi-modal problems, which can contribute to the construction of the general DEFI model.