Few-shot Incremental Event Detection Method Based On Knowledge

Event detection is used to quickly detect events from unstructured text and plays an important role in several natural language processing downstream tasks.Most existing methods only support detection of predefined event types,and adding new detection types requires retraining the model from scratch.With frequent new event types,retraining the model is a time-consuming,labor-intensive and difficult task to implement.Using incremental learning methods can solve the above problems to some extent.But traditional incremental learning still requires a large amount of high-quality data for new types of events.And there is a potential risk of catastrophic forgetting when training new classes.To address the above problems,we introduce few-shot learning in incremental learning to solve the problems of large amount of data for new types of events in incremental learning and catastrophic forgetting when training new classes.We add external knowledge,sample retention and hybrid distillation to achieve the goal of quickly adding new event detection capability to the model.Based on the above reasons,we mainly study the following two aspects:(1)We define a new task,few-shot incremental event detection,which focuses on learning to detect a new event class with limited data,while retaining the ability to detect old classes to the extent possible.We created a benchmark dataset IFSED for the few-shot incremental event detection task based on Few Event.(2)We use the classifier approach for model construction,for which a few-shot incremental event detection model with external knowledge(IFSED-K)is constructed.And we also use a meta-learning approach for model construction,for which a few-shot incremental event detection model with external knowledge and prototype network(IFSED-KP)is constructed.Experimental results show that our approach has a higher F1-score than baseline methods and is more stable.(3)We propose a method for few-shot incremental event detection based on external knowledge for sample retention(IFSED-K＿E)based on the IFSED-K model.We use a combination of external knowledge and training samples for sample retention,which helps the model to select more representative samples.The experimental results show that the method has significant improvement over other methods.