[Paper Review] One for All: Neural Joint Modeling of Entities and Events

Trung Minh Nguyen, Thien Huu Nguyen|arXiv (Cornell University)|Dec 1, 2018

Topic Modeling36 references6 citations

TL;DR

This paper proposes a unified deep learning model that jointly performs entity mention detection, event trigger identification, and argument role prediction using shared bidirectional RNN hidden representations. By eliminating error propagation from separate pipelines and leveraging end-to-end learning, the model achieves state-of-the-art F1 scores of 81.2 (EMD), 69.8 (ED), and 52.1 (ARP) on the ACE 2005 dataset without relying on handcrafted features.

ABSTRACT

The previous work for event extraction has mainly focused on the predictions for event triggers and argument roles, treating entity mentions as being provided by human annotators. This is unrealistic as entity mentions are usually predicted by some existing toolkits whose errors might be propagated to the event trigger and argument role recognition. Few of the recent work has addressed this problem by jointly predicting entity mentions, event triggers and arguments. However, such work is limited to using discrete engineering features to represent contextual information for the individual tasks and their interactions. In this work, we propose a novel model to jointly perform predictions for entity mentions, event triggers and arguments based on the shared hidden representations from deep learning. The experiments demonstrate the benefits of the proposed method, leading to the state-of-the-art performance for event extraction.

Motivation & Objective

• To address error propagation in traditional pipelined event extraction systems where entity mention detection errors affect downstream trigger and argument prediction.
• To enable knowledge sharing and interaction among entity mention detection, event trigger detection, and argument role prediction via shared hidden representations.
• To develop an end-to-end neural model that avoids reliance on manually engineered features like dependency paths or gazetteers.
• To improve performance on event extraction by jointly modeling all three subtasks using deep learning instead of discrete feature engineering.

Proposed method

• Uses a bidirectional RNN to generate shared hidden representations for all words in a sentence, capturing contextual dependencies across the entire sequence.
• Makes predictions for entity mention detection, event trigger detection, and argument role prediction directly from the shared hidden states.
• Employs a joint CRF layer to model label dependencies across the three subtasks, improving consistency and reducing error propagation.
• Trains the entire model end-to-end using a joint loss function that combines all three subtask objectives.
• Eliminates reliance on external features by learning contextual representations directly from word embeddings and the RNN architecture.
• Utilizes attention mechanisms to focus on relevant context for each prediction task, enhancing representation quality.

Experimental results

Research questions

• RQ1 Can joint modeling of entity mention detection, event trigger detection, and argument role prediction using shared deep representations outperform pipelined approaches?
• RQ2 To what extent can a single end-to-end deep learning model replace the need for handcrafted features in event extraction?
• RQ3 How does shared hidden representation learning improve performance across all three subtasks compared to isolated modeling?
• RQ4 What is the impact of removing external features on model performance, and can a pure neural approach still achieve state-of-the-art results?

Key findings

• The proposed end-to-end deep learning model achieves state-of-the-art F1 scores of 81.2 on entity mention detection, 69.8 on event trigger detection, and 52.1 on argument role prediction on the ACE 2005 dataset.
• Even without external features, the end-to-end model (End-to-end-DL) achieves comparable performance to state-of-the-art models that rely heavily on handcrafted features, outperforming Joint-Feature-Document by 1.9% on argument prediction.
• The model reduces error propagation by jointly learning all three subtasks, with trigger identification and argument classification benefiting from shared contextual representations.
• Error analysis shows that the main challenges lie in detecting rare triggers (e.g., 'intifada') and confusing similar roles (e.g., Place vs. Destination), indicating room for architectural improvements.
• The removal of external features leads to only a modest performance drop, demonstrating the robustness and generalization capability of the shared representation learning approach.