小样本增量事件检测是在样本数量有限的情况下,从类增量数据流中持续学习新事件类型,其中不同事件类型的样本在不同时间到达,同时最大程度地保留检测旧类的能力.然而,现有的工作仍不能解决灾难性遗忘问题.针对灾难性遗忘问题,本文提出了融合知识的原型矫正方法和自适应增量蒸馏.具体来说,本文对外部知识和训练样本分别进行编码和融合,绘制每个原型样本,以减小计算原型与实际原型之间的误差,对原型进行矫正.自适应增量蒸馏引用了元学习的思想,在蒸馏过程中通过增量数据,使学生模型提供教学反馈,更新教师模型的教学方式,使旧知识更好地传递到学生模型中.本文使用上述方法解决了灾难性遗忘问题.在小样本增量事件检测数据集(Few-Shot Incremental Event Detection Task based on the Few Event Dataset,IFSED)上的实验结果表明,本文的方法在新类上的表现,F1分数提高1.98％,在所有训练轮次可见的旧类上的表现,F1分数平均提高4.24％,在小样本增量事件检测任务上达了最佳性能.
Few-shot Incremental Event Detection Based on Prototype Correction and Adaptive Incremental Distillation
Few-shot incremental event detection involves continuously learning new event types from class-incremental data streams under limited sample availability,where samples of different event types arrive at different times,while maximally preserving the ability to detect old classes. However,existing works still fail to address the issue of catastrophic forgetting. To tackle catastrophic forgetting,we propose a knowledge-integrated prototypical rectification method and adaptive incremental distillation. Specifically,we encode and integrate external knowledge and training samples separately,mapping each prototype sample to minimize the error between the computed prototype and the actual prototype,thereby rectifying the prototype. Adaptive incremental distillation leverages the concept of meta-learning,allowing the student model to provide instructional feedback through incremental data during the distillation process,updating the teaching methods of the teacher model to better transfer old knowledge to the student model. Using the aforementioned methods,we have addressed the issue of catastrophic forgetting. Experimental results on the Few-shot Incremental Event Detection Task based on the Few Event Dataset( IFSED) show that our method improves performance on new classes by 1.98％ in F1 score,and by an average of 4.24％ on all old classes visible across all training rounds,achieving optimal performance in the task of few-shot incremental event detection.
NETL
NSTL
万方数据
