摘要
针对缺乏高质量标注数据的高鲁棒性要求领域事件检测任务,提出了一个集成多种鲁棒学习方法与模型结构的模型。首先,使用数据清洗和数据增强技术,缓解了训练集明显噪声与类别严重不均衡问题;然后,尝试了对抗学习方法和多种鲁棒损失函数,并在带噪声数据上进行鲁棒学习;最后,集成多种预训练模型与pipeline、joint 2种模型结构,得到鲁棒的事件检测结果。检测结果表明,该模型在训练集质量不高的情况下具有较强的鲁棒性和普适性。
Aimed at the domain event detection tasks requiring high robustness and lack of high-quali-ty annotated data,a model integrated multiple robust methods and model structures is proposed.First-ly,data cleaning and data augmentation techniques are used to alleviate the problem of significant noise and severe class imbalance in the training set.Then,adversarial learning methods and multiple robust loss functions are attempted to robust learning on noisy data.Finally,integrating multiple pre-training models with pipeline and joint model structures,robust event detection result is obtained.Re-sult shows that the proposed model has strong robustness and high universality under the condition of low-quality training data.
作者
唐旻骥
王振宇
丁效
杨重阳
曾屹荣
TANG Minji;WANG Zhenyu;DING Xiao;YANG Chongyang;ZENG Yirong(Research Center for Social Computing and Information Retrieval,Harbin Institute of Technology,Harbin 150001,China;National Key Laboratory of Information Systems Engineering,Nanjing 210023,China)
出处
《指挥信息系统与技术》
2024年第4期24-29,共6页
Command Information System and Technology
基金
国家自然科学基金(62176079和61976073)
科技部科技创新重大项目(2018AAA0101901)
信息系统工程全国重点实验室开放基金(052022077)资助项目。
关键词
领域事件检测
鲁棒性
数据处理
模型集成
domain event detection
robustness
data processing
model integration
