Abstract
Event Extraction (EE) is a challenging task in Natural Language Processing that requires identifying event triggers and their associated arguments. Traditional classification-based approaches often fail to capture the semantic meaning of argument roles and the dependencies between them. This paper discusses MQAEE, a framework that reformulates event extraction as a multi-turn question-answering task. By leveraging Machine Reading Comprehension (MRC) and history answer embeddings, MQAEE achieves state-of-the-art performance and demonstrates superior generalization to unseen event types.
Introduction
In the field of Information Extraction, Event Extraction (EE) aims to identify occurrences of specific event types and extract the entities involved in various roles. For example, in a "Movement:Transport" event, the system must identify the trigger (e.g., "left"), the artifact being moved (e.g., "Saddam's family"), and the origin (e.g., "that city").
Traditional systems typically treat EE as a classification problem, where event types and argument roles are represented as discrete labels. This paradigm suffers from "semantic blindness," as models do not utilize the inherent meaning of role names, and "dependency gaps," where the relationship between different arguments is ignored.
Methodology: The MQAEE Framework
The MQAEE (Multi-turn Question Answering for Event Extraction) framework addresses these issues by reframing the task into a conversational dialogue. The process is divided into three main stages:
- Trigger Identification: The model is asked an extractive question to find the trigger word in the text.
- Trigger Classification: A binary question confirms the specific event type associated with the identified trigger.
- Argument Extraction: This core stage involves a multi-turn dialogue where the model asks specific questions for each potential argument role.
To maintain context across turns, the researchers introduced History Answer Embeddings. These embeddings mark tokens that have already been identified as answers in previous turns, allowing the model to model "if-then" relationships between arguments (e.g., identifying a 'Buyer' makes it easier to locate the 'Seller').
Experimental Results
MQAEE was evaluated on the ACE 2005 dataset, a standard benchmark for event extraction. The model, built on BERT-Large, achieved a state-of-the-art F1 score of 53.4% for argument extraction, representing a 2.0% absolute improvement over previous benchmarks like DYGIE++.
Impact of Multi-turn Context
A key finding of the study was the importance of the multi-turn interaction. Ablation studies showed that removing the history context led to a significant performance drop.
As illustrated in the figure above, the argument classification F1 score generally increases as more history turns are included, confirming that the model effectively utilizes information from previous turns to improve current extractions.
Generalization Capabilities
One of the most significant advantages of MQAEE is its ability to generalize. By converting labels into natural language questions (e.g., "Who is the victim?"), the model can leverage the semantic knowledge encoded in pre-trained language models. In few-shot learning scenarios, MQAEE significantly outperformed traditional classifiers, maintaining competitive performance even with as few as five training samples for new event types.
Conclusion
MQAEE provides a more intuitive and semantically rich approach to event extraction. By treating the task as a multi-turn conversation, the model captures the complex dependencies inherent in real-world events. Future work may explore the use of Reinforcement Learning to optimize question ordering and address potential error propagation in the pipeline.
Summary updated