[Paper Review] Joint Event and Temporal Relation Extraction with Shared Representations and Structured Prediction

Rujun Han, Ning Qiang+1 more|arXiv (Cornell University)|Sep 2, 2019

Topic Modeling33 references4 citations

TL;DR

This paper proposes a joint neural model for end-to-end event and temporal relation extraction using shared contextual representations and structured prediction via integer linear programming (ILP). By training event and relation prediction jointly with structural constraints (consistency and transitivity), the method reduces error propagation and improves F1 scores by 10% on TB-Dense and 6.8% on MATRES compared to prior state-of-the-art systems.

ABSTRACT

We propose a joint event and temporal relation extraction model with shared representation learning and structured prediction. The proposed method has two advantages over existing work. First, it improves event representation by allowing the event and relation modules to share the same contextualized embeddings and neural representation learner. Second, it avoids error propagation in the conventional pipeline systems by leveraging structured inference and learning methods to assign both the event labels and the temporal relation labels jointly. Experiments show that the proposed method can improve both event extraction and temporal relation extraction over state-of-the-art systems, with the end-to-end F1 improved by 10% and 6.8% on two benchmark datasets respectively.

Motivation & Objective

• To address error propagation in pipeline systems that separately perform event extraction and temporal relation classification.
• To improve event representation by sharing contextualized embeddings and neural feature learners between event and relation modules.
• To enable end-to-end prediction of events and temporal relations using structured inference with global constraints.
• To validate that joint learning with structural constraints enhances both event and relation extraction performance.
• To analyze failure modes and identify limitations in detecting VAGUE relations and low-confidence events.

Proposed method

• Uses BERT-based contextualized embeddings as shared input representations for both event and relation modules.
• Employs a shared neural representation learner to extract deep features from the same input for both tasks.
• Applies structured prediction via integer linear programming (ILP) to enforce global constraints: event-relation consistency and temporal transitivity.
• Implements structural support vector machines (SSVM) for end-to-end joint training with global label consistency.
• Enforces that relations between non-events or event-non-event pairs must be NONE, and no temporal loops are allowed.
• Uses confidence thresholds to filter low-scoring event predictions and applies joint inference to correct misclassified events.

Experimental results

Research questions

• RQ1 Can joint learning of event and temporal relation extraction reduce error propagation from pipeline systems?
• RQ2 Does sharing representations between event and relation modules improve performance on both tasks?
• RQ3 Can structured prediction with global constraints (consistency and transitivity) enhance the accuracy of end-to-end extraction?
• RQ4 How do different structural constraints (e.g., consistency vs. transitivity) contribute to performance gains?
• RQ5 What are the primary failure modes in joint event and relation prediction, especially for VAGUE relations and low-confidence events?

Key findings

• The proposed joint model improves end-to-end F1 by 10% on the TB-Dense dataset compared to the CAEVO baseline.
• On the MATRES dataset, the model achieves a 6.8% improvement in end-to-end F1 over the CogCompTime baseline.
• The event-relation consistency constraint contributes significantly, improving F1 by 0.9% on TB-Dense and 1.0% on MATRES.
• The transitivity constraint provides minimal improvement, likely due to BERT’s contextual modeling and the presence of NONE relations.
• Type 1 errors (false non-events) and Type 2 errors (false positive relations on non-events) are the most frequent failure modes, indicating limitations in event detection confidence.
• VAGUE relations are a major challenge, often confused with other temporal relations, especially when combined with NONE pairs in end-to-end settings.