Abstract
Event extraction (EE) involves identifying event triggers and their associated arguments. Traditional methods often struggle with argument role ambiguity because they treat extraction as a simultaneous or fixed-sequence task, ignoring inter-argument correlations. This paper proposes a novel framework that reformulates EE as a multi-turn, task-oriented dialogue. By employing reinforcement learning (RL) to optimize the extraction order, the model leverages previously identified arguments to resolve the roles of more complex ones. Evaluated on the ACE 2005 benchmark, the approach demonstrates significant improvements over state-of-the-art methods.
Introduction: Contextual Dependencies in Event Extraction
The primary goal of Event Extraction is to identify "triggers"—words that signify an event's occurrence—and their "arguments," which represent participants, locations, or attributes. A significant challenge in this field is that an argument's role is often context-dependent. For instance, the word "troops" might function as an attacker in one sentence and a target in another, depending on the presence of other entities like "weapons" or "buildings."
Most existing models fail to capture these dependencies effectively. The authors argue that identifying "easier" arguments first can provide the necessary context to disambiguate more difficult ones.
Proposed Framework
The proposed system consists of three integrated components designed to capture both global semantics and local context through an iterative process.
1. Double Perspective Event Representation
The model constructs a dual-view representation of the input text:
- Lexicon-based Graph Attention Network (GAT): This component treats words as nodes in a graph to capture long-range dependencies and global semantic structures.
- Event-based BERT: This utilizes contextual embeddings specifically fine-tuned for the nuances of event extraction.
2. Ranked Argument Extraction via Dialogue
The core innovation is modeling the extraction process as a conversation between two agents. Agent A (The Questioner) generates queries based on the event schema (e.g., "What is the killing instrument?"), while Agent B (The Answerer) identifies the corresponding text span.
A Reinforcement Learning (RL) policy is used to determine the optimal sequence of questions. By learning which arguments are most informative or easiest to identify, the RL agent guides the dialogue to build a robust context for subsequent turns.
3. Incremental Event Learning
As the dialogue progresses, newly extracted arguments are fed back into the GAT as "pseudo-labels." This updates the sentence representation in real-time, allowing the model's understanding to evolve dynamically with each turn.
Experimental Evaluation
The framework was evaluated using the ACE 2005 dataset across English, Arabic, and Chinese languages. The dataset includes 33 event subtypes and 36 argument roles.
The "Full Model" achieved an F1-score of 61.42% on the Argument Role Classification (ARC) task, significantly outperforming previous machine reading comprehension (MRC) baselines which typically scored between 53% and 59%. Ablation studies revealed that the Reinforcement Learning (RLD) module was the most critical component; its removal led to the sharpest decline in performance, confirming that the optimized extraction order is vital for success.
Discussion and Future Directions
While the dialogue-based approach is highly effective, it introduces increased computational complexity. The training time for the full model was approximately 38.1 hours, compared to roughly 20-24 hours for simpler baselines. Additionally, the model occasionally struggles with complex sentences where arguments are widely scattered, potentially losing track of distant relations in the dialogue history.
Future research will focus on integrating commonsense knowledge to further resolve ambiguities and optimizing the model's efficiency to reduce the computational overhead of the multi-turn dialogue process.
Summary updated