# Document-Level Event Extraction via Human-Like Reading Process

**Authors:** Shiyao Cui, Xin Cong, Bowen Yu, Tingwen Liu, Yucheng Wang, Jinqiao Shi  
**arXiv:** [2202.03092](https://arxiv.org/abs/2202.03092)  
**Date:** 2022-02-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.03092v1/

## Key points

- Challenges of Document-level Event Extraction (DEE): scattering-arguments and multi-events
- Proposed method: HRE (Human Reading inspired Extractor for Document Events)
- Two iterative stages: rough reading and elaborate reading
- Rough reading stage: scan document to detect occurrence of events
- Elaborate reading stage: hierarchical approach to extract specific event arguments
- Addressing scattering-arguments problem by systematically identifying and connecting relevant information
- Multi-round rough reading to uncover undetected events, addressing multi-events challenge
- Experimental results show HRE outperforms previous methods in DEE tasks
- HRE leverages insights from human cognitive processes during reading for improved accuracy and efficiency in information extraction

## Summary

The paper titled "Document-Level Event Extraction via Human-Like Reading Process" addresses the challenges of Document-level Event Extraction (DEE), namely scattering-arguments and multi-events. Scattering-arguments refers to the situation where arguments of one event record are spread across different sentences in a document, while multi-events indicate that a document can contain multiple event records simultaneously. To tackle these challenges, the authors propose a method called HRE (Human Reading inspired Extractor for Document Events) that decomposes DEE into two iterative stages: rough reading and elaborate reading. In the rough reading stage, the document is scanned to detect the occurrence of events. This initial pass helps identify potential events within the text. The second stage, elaborate reading, focuses on extracting specific event arguments. Here, a hierarchical approach is employed, working from sentences to characters to locate arguments across sentences. This effectively addresses the scattering-arguments problem by systematically identifying and connecting relevant information. Additionally, rough reading is explored in a multi-round manner to uncover undetected events. By iteratively revisiting the document, potentially missed events can be discovered and included in the extraction process. This approach effectively handles the challenge of multi-events. Experimental results demonstrate that HRE outperforms previous competitive methods in document-level event extraction tasks by leveraging insights from human cognitive processes during reading to improve information extraction accuracy and efficiency. Overall, this paper presents an innovative approach that addresses DEE challenges through a two-stage iterative process inspired by human reading patterns which highlights its superiority over existing methods in extracting events from documents at a comprehensive level.