ODEE: A One-Stage Object Detection Framework for Overlapping and Nested Event Extraction

Jinzhong Ning, Zhihao Yang, Zhizheng Wang, Yuanyuan Sun, Hongfei LinIJCAI 2023 · DOI

TL;DR

This paper introduces ODEE, a novel one-stage framework that treats event extraction as an object detection problem. By reformulating event extraction as identifying and grouping token pair vertices, ODEE effectively handles complex overlapping and nested events, achieving state-of-the-art performance while improving computational efficiency.

What Problem Does This Paper Address?

Traditional event extraction (EE) methods often assume events are "flat," meaning they don't overlap or nest within each other. However, real-world text frequently contains events that are intertwined. For instance, one event's trigger might be another event's argument, or multiple events might share parts of their triggers or arguments. Existing methods struggle with these "overlapping" and "nested" events, often relying on multi-stage pipelines that can propagate errors. While some recent approaches like table-filling methods have improved this, they can be computationally expensive and suffer from imbalanced training data. This paper aims to address these limitations by proposing a unified, efficient, and accurate method for overlapping and nested event extraction.

Methods

The core innovation of ODEE is its reformulation of event extraction as an object detection task. Instead of traditional sequence labeling or table-filling, ODEE treats event triggers and arguments as "objects" to be detected within a "table" of token pair representations.

Here's a breakdown of the key techniques:

  • Object Detection Analogy: The problem is framed as identifying and grouping four vertices (upper-left, upper-right, lower-left, lower-right) that define the bounding box of a trigger-argument pair in a token-pair representation table. This allows the model to directly predict the spatial extent of events.
  • Unified Encoder: To better capture the relationship between event types and the input sentence, event type names are verbalized into natural language text. This text is then concatenated with the input sentence and fed into a pre-trained BERT model. This allows for a unified semantic representation where event types and sentence context interact.
  • Span-Aware Auxiliary Tasks: To ensure the model fully utilizes span information, ODEE incorporates two auxiliary tasks: predicting the type and span of trigger words, and predicting the type and span of argument entities. These tasks act as guidance during training.
  • Negative Sampling for Table Cells: To combat the imbalance between positive (actual event components) and negative (non-event components) table cells, a negative sampling strategy is employed. This preserves all positive tokens and randomly samples a fixed number of negative tokens, significantly improving training efficiency and reducing computational cost.
  • Relation-Specific Vertex Regressor: A specialized regressor is used to predict the probability of each token pair belonging to a specific vertex tag type (UL, UR, LL, LR) for a given relation type.

Datasets

The paper evaluates ODEE on three benchmark datasets:

  • FewFC: A Chinese financial event extraction dataset, notable for containing a significant proportion of overlapping events.
  • Genia11 & Genia13: Biomedical event extraction datasets that are rich in nested events.

The datasets are split into training, development, and testing sets following standard practice.

Key Insights and Ideas

  • Event Extraction as Object Detection: The most significant insight is the successful adaptation of object detection principles from computer vision to the NLP task of event extraction. This allows for a more direct and holistic approach to identifying complex event structures.
  • Vertex-Based Tagging: Using the four vertices of a token-pair region to define event spans is a novel way to capture the spatial relationships between triggers and arguments, leading to better utilization of span information.
  • Unified Representation is Key: Combining event type verbalizations with sentence embeddings via BERT creates a richer, unified representation that improves the model's understanding of event context.
  • Negative Sampling for Efficiency and Balance: The negative sampling strategy is crucial for making the table-filling approach computationally feasible and addressing the inherent data imbalance, leading to better performance.

Key Results

  • State-of-the-Art Performance: ODEE achieves state-of-the-art F1 scores across all three benchmark datasets (FewFC, Genia11, Genia13) for overlapping and nested event extraction.
  • Significant Improvements: Compared to sequence labeling methods, ODEE shows substantial gains in recall and F1 scores. It also outperforms multi-stage methods by reducing error propagation.
  • Efficiency Gains: ODEE is more computationally efficient than existing state-of-the-art methods like CasEE and OneEE, boasting fewer parameters, faster inference speeds, and lower GPU memory usage. For example, on the FewFC dataset, ODEE achieves 23.8 sentences/second inference speed, significantly faster than CasEE (12.5 sent/s) and OneEE (14.7 sent/s).

Limitations and Future Work

  • Ablation Study Findings: The paper highlights that both auxiliary tasks (trigger type/span prediction and argument type/span prediction) are crucial for performance. Removing either leads to a drop in F1 scores. The negative sampling method is also shown to be vital for alleviating label imbalance and improving efficiency.
  • Future Work: While not explicitly detailed, the paper's success in unifying event types and sentence representations, along with its efficient one-stage approach, suggests that further exploration of pre-trained language models and object detection techniques for complex IE tasks is a promising direction. The authors also make their code publicly available, facilitating future research.

Summary updated

BibTeX
@inproceedings{ning2023odee,
  author = {Jinzhong Ning and Zhihao Yang and Zhizheng Wang and Yuanyuan Sun and Hongfei Lin},
  title = {{ODEE: A One-Stage Object Detection Framework for Overlapping and Nested Event Extraction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/574},
}
Figure