Research On Response Safety Of Online Learning Chatbot

With the rapid development of artificial intelligence technology,more and more relevant technical achievements have begun to come out of the laboratory and enter the market,including chatbots and personal assistants such as Tay,Xiao-Ice,Alex and Siri.While admiring the many conveniences to life brought by AI products,people also began to worry about the safety of these products.Facts have proved that these concerns are not superfluous,many security issues have emerged during the productization process of AI technologies,including frequent attacks against vulnerabilities in chatbot's online learning interface: Meaning that hackers or malicious users use such vulnerabilities to "teach" the chatbot offensive responses that are in violation of local laws and regulations.More often than not,chatbot products are forced to be removed due to improper responses,causing great losses to related companies.Due to it is impossible to determine the specific time when the model learned the offensive responses,it is difficult to accurately find the uncontaminated version.In this case,the product can only be rolled back to an early version,resulting in the loss of many valuable contents already learned,which often leads to great difficulty in the subsequent rectification work.In view of the above,those in industrial and academic circles are anxious to find solutions that can ensure the reply safety of chatbot with online learning interface,which is what this thesis is trying to address.The works and contributions of this thesis are as follows:1.We proposed a safe response framework for online learning chatbot.The framework includes an input sentence-aware censor model,an online active learning algorithm for censor model and an utterance purification algorithm based on reinforcement learning.In addition,the above framework also boasts strong flexibility by integrating various algorithm processes in a loosely coupled manner.This thesis also introduces a chatbot application example established using the above framework.2.We proposed an input sentences-aware censor model.Unsafe responses of chatbots can be divided into the following three categories: There are explicit profane words in response sentences;There are no explicit profane words in the response sentence,but the semantics of the sentence are offensive;There are no explicit offensive words or semantics in the response sentence,but it is offensive if the context of the input is considered.It can be seen that in addition to detecting the first two kinds of unsafe responses,the censor also needs to take into account the semantics of user input sentences.Therefore,this thesis proposes an encoder-classifier architecture that introduces the semantics of user input sentences into the classification process.Experimental results show that this architecture can improve the accuracy of unsafe responses detection and the alleviate the problem of long-term dependence caused by the introduction of the user input sentences.3.We proposed an online active learning algorithm for continuous learning of censor models.In response to the rapid evolution and increase of offensive languages,this thesis introduces an online learning interface for user reporting.However,this interface is also prone to being used by hackers or malicious users to interfere with the classification workflow of the censor model.To address this problem,the author proposes to add an topic model-based active learning algorithm which can not only preferentially select samples with low confidence from user feedbacks for manual annotation and confirmation,but also enhance the feature extraction capability of the censor by learning form large amounts of feedback data yet to be manually confirmed.Experiments on multiple datasets prove that this algorithm can help reduce the cost of manual annotation while ensuring the safety of the feedback interface.4.We proposed a speech purification algorithm based on reinforcement learning.When the chatbot model has been corrupted,using this algorithm allows the chatbot to forget through reinforcement learning the unsafe responses learned without having to roll back to a previous version.At the same time,by integrating one-shot learning,it's possible to perform fast speech purification without forgetting the basic grammar learned before.Experiments prove that this algorithm helps reduce the probability of chatbot generating unsafe responses,while integration of one-shot learning also helps to improve the training speed and reduce the impact on response sentence fluency at the same time.