Trigger is Not Sufficient: Exploiting Frame-aware Knowledge for Implicit Event Argument Extraction

Kaiwen Wei, Xian Sun, Zequn Zhang, Jingyuan Zhang, Guo Zhi, Li JinACL 2021

TL;DR

This paper introduces FEAE, a novel framework for Implicit Event Argument Extraction (IEAE) that leverages "frame-aware" knowledge to improve reasoning over long-range and implicit dependencies. By employing a curriculum knowledge distillation strategy with a teacher-student model, FEAE achieves state-of-the-art performance on the RAMS dataset without requiring extra input tools during inference.

What Problem Does This Paper Address?

Extracting arguments related to an event (like who did what, where, and with what) is a core task in Natural Language Processing. However, existing methods often struggle when arguments are not explicitly stated in the same sentence as the event trigger or when they are separated by many sentences. This is particularly challenging for "Implicit Event Argument Extraction" (IEAE), where arguments can span multiple sentences and require deeper reasoning. Current approaches primarily focus on direct relationships between event triggers and arguments, lacking the ability to infer these implicit connections effectively. This limitation hinders their performance on complex, real-world scenarios where information is distributed and context-dependent.

Methods

The core of FEAE is a Frame-aware Event Argument Extraction framework. It builds upon Machine Reading Comprehension (MRC) models, which frame argument extraction as a question-answering task. The novelty lies in how FEAE incorporates "frame-level" knowledge, essentially using other related arguments within an event's context as clues to help identify the target argument.

The framework employs a teacher-student knowledge distillation strategy.

  • Teacher Model: A well-informed teacher model is trained using "oracle knowledge" (ground-truth information) to understand event frames and reasoning.
  • Student Model: A student model, which does not have access to this oracle knowledge during inference, learns to mimic the teacher's behavior. This allows the student to perform reasoning without needing explicit external tools.

To further bridge the gap between training with perfect knowledge and testing with imperfect real-world data, FEAE introduces a curriculum distillation strategy. This strategy gradually increases the learning complexity for the student model, starting with more complete oracle information and progressively reducing it, mimicking a curriculum that eases the learning process. The framework also supports a multi-teacher setup, where multiple teachers, trained with different subsets of related arguments, provide diverse guidance to the student.

Datasets

The primary dataset used for training and evaluation is the RAMS dataset. This dataset is annotated with 139 event types and 65 argument roles, featuring contexts of around five sentences. The paper notes that RAMS contains 7329 training, 924 development, and 871 test instances.

Key Insights and Ideas

  • Frame-level reasoning is crucial for IEAE: Simply looking at direct trigger-argument relationships is insufficient. Utilizing other arguments within an event's frame provides vital clues for inferring implicit arguments and handling long-range dependencies.
  • Knowledge distillation can bridge the train-test gap: The disparity between having perfect oracle knowledge during training and facing imperfect inputs during testing is a major challenge. Knowledge distillation effectively transfers the reasoning capabilities of a teacher model to a student model that operates without such privileged information.
  • Curriculum learning enhances robustness: Gradually increasing the complexity of the learning task for the student model, by varying the amount of provided related arguments, leads to better adaptation to real-world scenarios and improved performance.
  • Multi-teacher guidance offers diverse perspectives: Training multiple teacher models with different views of the event frame can provide richer and more robust guidance to the student model.

Key Results

FEAE achieves state-of-the-art performance on the RAMS dataset for both argument identification and classification.

  • The proposed FEAE model outperforms strong baselines by a significant margin, achieving an F1 score of 53.49 for argument identification and 47.40 for classification.
  • The framework demonstrates a substantial improvement in handling long-range dependencies, showing a gain of more than 4 times in F1 score for arguments at distances of ±2 sentences compared to a baseline student model.
  • Ablation studies confirm the effectiveness of each component: knowledge distillation contributes significantly to argument classification, and the curriculum strategy helps bridge the train-test gap.

Limitations and Future Work

  • Limitations: The paper acknowledges that while FEAE performs well, the train-test disparity can still affect performance. The teacher model relies on oracle knowledge, which is not available in real-world applications. The performance breakdown by distance shows that while FEAE mitigates the issue, performance still drops for non-local arguments.
  • Future Work: The authors suggest further exploring the BERT attention mechanism to better understand the reasoning process. They also propose investigating more advanced curriculum learning strategies and exploring multi-turn event extraction scenarios. The potential for applying FEAE to other NLP tasks that involve complex reasoning and implicit information is also an avenue for future research.

Summary updated

BibTeX
@inproceedings{wei2021trigger,
  author = {Kaiwen Wei and Xian Sun and Zequn Zhang and Jingyuan Zhang and Guo Zhi and Li Jin},
  title = {{Trigger is Not Sufficient: Exploiting Frame-aware Knowledge for Implicit Event Argument Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure