Research On Deep Slot Filling In Noisy And Imbalanced Data Scenarios

Slot filling,as a key module of a spoken language understanding system,aims to extract entity slot values in input text and fill them into predefined slots.With the development of deep learning technology,the data-driven supervised learning method based on sequence annotation framework has achieved excellent results in slot filling tasks.However,the performance of existing models suffers severely in real-world scenarios with noisy data and imbalanced data distributions.Therefore,how to utilize deep neural networks to solve the challenges of slot filling tasks in noisy and imbalanced data scenarios has become a hot research topic.This paper focuses on the above two challenges and presents the following contributions:(1)To address the lack of evaluation datasets for slot filling with natural noise,this paper constructs a noise-robust slot filling evaluation test set that contains five types of real-world noise scenarios,which can be used to test the robustness of models in various noisy data scenarios.(2)To address the issue of input noise causing model disturbance in noisy data scenarios,this paper designs a generic two-stage enhancement robust training framework that integrates various noise-robust training methods.Based on extensive experiments on the noisy test set proposed in this paper,this paper comprehensively analyzes the noise types and robust training methods to promote research in this direction.(3)Slot filling with scarce labels can be addressed using a self-training framework.However,existing self-training models overlook the problem of imbalanced slots,resulting in the performance degradation of minorityclass slots during iteration.This paper proposes a novel self-training model for imbalanced slot filling that aims to learn unbiased margins among slot classes,while reducing potential slot confusion,and adaptively sampling pseudo-labeled data to balance the slot distribution in the training set.This model achieves an improvement in Macro F1 score on two datasets and establishes the state-of-the-art performance for semi-supervised slot filling tasks.In response to the real-world challenges faced by slot filling tasks,this paper proposes two new methods from two perspectives and demonstrates their effectiveness on multiple datasets.Slot filling tasks in noisy and imbalanced data scenarios still face many challenges,and further research is needed to address them.