LwF4IEE: An Incremental Learning Method for Interactive Event Extraction

• Published in: Proceedings (International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery. Online) pp. 104 - 113
• Main Authors: Duan, Jiashun, Zhang, Xin, Xu, Chi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.10.2022
• Subjects: candidate recommendation; Data mining; event extraction; incremental learning; interactive event extraction; Knowledge discovery; Learning systems; Manuals; Predictive models; Task analysis; Training data
• ISSN: 2833-8898
• DOI: 10.1109/CyberC55534.2022.00026

Albeit great progress has been witnessed in event extraction, the accuracies achieved up to now by various automatic models still can not meet the performance requirements of some special applications such as disaster monitoring and rescue. It motivates us to introduce a new human-in-loop extraction mode called interactive event extraction (IEE), which works iteratively. Each iteration consists of three main steps: "model recommending candidate results → manual selecting and correcting → model re-training and updating". For candidate recommendation, we build an MRC (Machine Reading Comprehension)-based model that can output several most likely candidate elements, i.e., candidate triggers and arguments, by confidence evaluation. For model re-training and updating, we proposed an incremental learning method named LwF4IEE (Learning without Forgetting for IEE), which employs manual selected and corrected results as hard label and prediction of original model as soft label to avoid catastrophic forgetting. We conduct extensive experiments on datasets constructed from real-world Chinese texts. The results show that when setting the number of candidates to be 5, recalls of triggers and arguments reach 93.80% and 90.58% respectively, which is 11.51% and 11.33% higher compared with the basic MRC-based automatic extraction model. Moreover, LwF4IEE increases the recall of triggers by 2.71% on specific event types and only decreases by 0.24% on other types, achieving the purpose of learning without forgetting.