Research On Robustness Methods In Natural Languag Understanding And Generation

Artificial intelligence technology has been widely used in various fields in recent years,showing excellent performance,and is expected to have a significant positive impact on human production patterns and lifestyles.Robustness is an important aspect of artificial intelligence technology,which is mainly used to describe whether the model can still maintain performance in the face of complex and variable or noisy input data.The level of robustness directly determines the generalization ability and application potential of models and systems.In the research field,many models can only show good results on a specific standard dataset,but it is difficult to replicate the same excellent performance on other datasets or real data,which is because the model is not sufficiently robust for the changes in the data.In real-world application scenarios,the model has to face more complex language application methods,and the data to be processed contains more complex changes.Once the robustness is lacking,the performance of the model in real applications will be greatly compromised.In domains with high social impact,such as navigation or medical diagnosis,a less robust model can even lead to catastrophic errors.Therefore,in order to ensure the practical application value of the model,it is necessary to conduct in-depth research on the robustness of the model.The content of robustness research usually changes with the task content performed by the model.In the field of natural language processing,tasks can be divided into two categories:natural language understanding and natural language generation.Therefore,this paper studies the robustness methods in natural language understanding and generation.From the perspective of representation learning,a number of novel and effective robustness methods are proposed.The main work of this paper is as follows.The robustness method of out of domain detection in natural language understanding of dialogue system is studied.The main contributions include:Ⅰ Aiming at the problem of out of domain detection in the scene with few supervised samples,a generative classifier based on Gaussian discriminant analysis is proposed to solve the over confidence problem in the out of domain detection method of deep learning;An implicit data augmentation method based on adversarial attack is proposed to alleviate the lack of training data.Experiments show that these methods are more robust to data imbalance,few samples and semantic overlap.ⅡA self-supervised contrastive learning framework for out of domain detection is proposed.The idea is to combine unsupervised and supervised learning,so as to optimize the semantic features inside and outside the domain at the same time.The framework can extract differentiated semantic representation from unlabeled data;In addition,an explicit and implicit data augmentation method is designed to improve the diversity and complexity of data.Experiments show that the performance of this method can approach the supervised out of domain detection.Ⅲ Furthermore,an out of domain detection method in unsupervised scene is proposed.This method adopts a contrastive learning training objective function based on discriminant representation analysis,which can learn the differentiated semantic intention representation through the data in domain.Experiments show that this method has better performance than the current SOTA method,and can improve the robustness of taskbased dialogue system.The robustness method of factual consistency in text generation is studied.The main contributions include:Ⅰ Aiming at the problem of fact consistency in text summary,a mechanism based on gradient tracking factual error under the adversarial attack mechanism is proposed to improve the interpretability and robustness of fact consistency evaluation.Experiments show that this method can effectively optimize the fact consistency evaluation,and provide available auxiliary information for manual judgment.Ⅱ Aiming at the problem of fact consistency in text generation of controlled data,based on the cross-table representation method and combined with the good properties of contrastive learning,a cell-level contrastive learning objective function is proposed.In addition,it also realizes the function of highlighting key contents.Experiments on standard datasets show the effectiveness of the proposed method and can have a positive impact on the robustness of text generation.