[Paper Review] Doc2EDAG: An End-to-End Document-level Framework for Chinese Financial Event Extraction

Shun Zheng, Wei Cao+2 more|arXiv (Cornell University)|Apr 16, 2019

Topic Modeling34 references4 citations

TL;DR

This paper proposes Doc2EDAG, an end-to-end document-level event extraction framework for Chinese financial announcements that models event extraction as entity-based directed acyclic graph (EDAG) generation. By leveraging document-level context encoding and a memory-augmented path-expanding mechanism, it outperforms state-of-the-art methods by large margins, especially on multi-event and argument-scattering scenarios, achieving up to 28.4 F1 improvement over the best baseline on the EO event type.

ABSTRACT

Most existing event extraction (EE) methods merely extract event arguments within the sentence scope. However, such sentence-level EE methods struggle to handle soaring amounts of documents from emerging applications, such as finance, legislation, health, etc., where event arguments always scatter across different sentences, and even multiple such event mentions frequently co-exist in the same document. To address these challenges, we propose a novel end-to-end model, Doc2EDAG, which can generate an entity-based directed acyclic graph to fulfill the document-level EE (DEE) effectively. Moreover, we reformalize a DEE task with the no-trigger-words design to ease the document-level event labeling. To demonstrate the effectiveness of Doc2EDAG, we build a large-scale real-world dataset consisting of Chinese financial announcements with the challenges mentioned above. Extensive experiments with comprehensive analyses illustrate the superiority of Doc2EDAG over state-of-the-art methods. Data and codes can be found at https://github.com/dolphin-zs/Doc2EDAG.

Motivation & Objective

• To address the limitations of sentence-level event extraction in handling argument-scattering and multi-event scenarios common in financial documents.
• To propose a document-level, end-to-end event extraction model that avoids reliance on trigger words and key-event sentence heuristics.
• To build a large-scale, real-world dataset of Chinese financial announcements annotated for document-level event extraction.
• To evaluate the effectiveness of end-to-end modeling with document-level context and memory-augmented path expansion on complex event structures.
• To demonstrate the superiority of the proposed no-trigger-words formalization for distant supervision-based labeling.

Proposed method

• Doc2EDAG formulates document-level event extraction as generating an entity-based directed acyclic graph (EDAG), transforming the table-filling task into sequential path-expanding sub-tasks.
• It employs a Transformer-based encoder to generate document-level entity representations, capturing long-range dependencies across sentences.
• A memory mechanism is introduced to maintain and update path states during EDAG construction, enabling effective tracking of partial event structures.
• Scheduled sampling is used during training to reduce the distribution shift between training and inference by gradually incorporating predicted entities into the input.
• The model is trained end-to-end to directly predict event tables from raw documents, without requiring trigger-word labeling or key-event sentence detection.
• A no-trigger-words formalization is introduced, treating DEE as direct table-filling, which simplifies labeling and improves alignment with the ultimate goal of event extraction.

Experimental results

Research questions

• RQ1 Can an end-to-end document-level event extraction model outperform two-stage approaches that rely on key-event sentence detection and heuristic argument completion?
• RQ2 How effective is the EDAG generation strategy with memory-augmented path expansion in handling argument-scattering and multi-event scenarios?
• RQ3 Does removing the need for trigger-word labeling improve robustness and performance in distant supervision-based DEE?
• RQ4 To what extent does document-level context encoding enhance the model’s ability to resolve cross-sentence arguments and multiple events?
• RQ5 How does the scheduled sampling strategy mitigate the discrepancy between training and inference in sequential path expansion?

Key findings

• Doc2EDAG achieves a 19.1 F1 score improvement over the best baseline (DCFEE-O) on the EF event type, demonstrating strong performance on argument-scattering cases.
• On the multi-event test set, Doc2EDAG improves by 17.7 F1 points over DCFEE-O, highlighting its effectiveness in handling complex, overlapping event structures.
• The ablation study shows that removing the memory mechanism causes the largest performance drop—over 10 F1 points on four event types—proving its critical role in path tracking.
• Scheduled sampling improves average F1 by 5 points, confirming its value in reducing training-inference distribution shift.
• Document-level entity encoding contributes 2.1 F1 points on average, showing that global context enhances entity representation quality.
• The use of higher negative class weight for path expansion reduces false positives and yields stable, measurable improvements across all event types.