Research On Knowledge-Grounded Non-Task-Oriented Conversational System

Letting machines chat with human beings in natural(human)language is one of the key targets of Artificial Intelligence(AI).To achieve this goal,it's necessary to develop intelligent Human-Machine Conversational Systems(CSs).CSs have a wide range of application scenarios,and can be divided into two groups according to different purposes,i.e.,task-oriented CSs and non-task-oriented CSs(a.k.a.Chatbot).Compared to the task-oriented CSs,the existing nontask-oriented CSs are far from satisfactory.Non-task-oriented CSs are often implemented within the Seq2seq framework,such that they are able to generate natural and fluent responses but suffer from a serious" safe response" problem,i.e.,they tend to generate uninformative responses.Some non-task-oriented CSs mitigate this problem by introducing external knowledge,and they usually are referred to as Knowledge-Grounded Non-Task-Oriented Human-Machine Conversational Systems(KGCs).Generally,KGCs need to solve three tasks:1)Conversational Contexts and Knowledge Encoding(CCKE),i.e.,encoding conversational contexts(dialogue history between systems and users)and knowledge into latent representations;2)Knowledge Selection(KS),i.e.,selecting the appropriate knowledge to be incorporated in the current response);and 3)Response Generation(RG),i.e.,generating responses according to the selected knowledge and conversational contexts).This dissertation has carried out three research studies,aiming at further improving the performance of KGCs via enhancing their performance in terms of two important tasks,KS and RG.The main innovation points of this dissertation can be summarized as the following three aspects:(1)Propose a dual knowledge interaction model for KS taskExisting studies usually select knowledge based on conversational contexts,and optimize this process via Maximum Likelihood Estimation(MLE).However,there are two drawbacks for existing studies.Firstly,they do not explicitly model knowledge tracking and shifting,i.e.,first tracking the knowledge previously selected in the given conversational context(knowledge tracking)and then inferring the knowledge to be selected based on the selected knowledge and the given conversational context.This process can better capture the interaction between the knowledge previously selected and the knowledge to be selected in order to further improve the performance of KS.Secondly,MLE cannot capture the manyto-many mapping between conversational contexts and knowledge.Specifically,given a conversational context,there are multiple appropriate pieces of knowledge that can be selected(one-to-many mapping);conversely,given a piece of knowledge,it could be used in multiple conversational contexts(many-to-one mapping).Unfortunately,for a given conversational context,only one piece of knowledge is annotated as ground truth in existing datasets because manual annotation is resource intensive.MLE only regards the annotated knowledge as the correct knowledge such that it cannot capture the many-to-many mapping.This dissertation proposes a dual knowledge interaction model,which explicitly models knowledge tracking and shifting.Besides,this dissertation further formulates knowledge tracking and knowledge shifting as dual tasks,and devises dual knowledge interaction learning scheme based on Dual Learning in order to better facilitate interaction between them.This learning scheme not only lets knowledge tracking and shifting teach each other,improve together and iterate until convergence in an unsupervised learning way,but also explores and rewards extra appropriate knowledge that is not annotated as ground truth in training sets,which helps to capture the above-mentioned many-to-many mapping.Experimental results show that the dual knowledge interaction model significantly outperforms state-of-the-art methods.(2)Propose a mixed-initiative knowledge selection model for KS taskNo previous studies have considered the mixed-initiative characteristics of KS to improve its performance.Mixed Initiative is an intrinsic feature of conversations where the user and the system can both take the initiative in driving conversational directions.In fact,KS also has the mixed-initiative nature,i.e.,KS can be divided into user-initiative KS and system-initiative KS.For the former,the system usually selects knowledge according to the current user utterance that contains new topics or questions posed by the user;for the latter,the system usually selects knowledge according to the previously selected knowledge.This dissertation proposes a mixed-initiative knowledge selection model,which explicitly distinguishes between user-initiative and system-initiative KS.Concretely,this dissertation introduces two knowledge selectors to model both of them separately,and design a novel initiative discriminator to discriminate the initiative type of KS at each conversational turn.A challenge for training this model is that there are no labels for indicating initiative in datasets.To tackle this challenge,based on assumptions,this dissertation devises an initiative-aware self-supervised learning scheme that helps the mixed-initiative knowledge selection model learn to discriminate the initiative type via an approximately equivalent self-supervised task.Experimental results show that the mixed-initiative knowledge selection model significantly outperforms state-of-the-art methods.(3)Propose a reference-aware model for RG taskExisting studies mainly generate responses via generation-based methods,i.e.,predicting tokens one by one from pre-defined vocabulary or additionally copying tokens one by one from external knowledge.The above generation-based methods are good at generating natural and fluent responses,but suffer from the problems of low efficiency in leveraging external knowledge and tending to break the complete semantic units from external knowledge,respectively.Some studies produce responses via extraction-based methods,i.e.,directly extracting a complete semantic unit from external knowledge as a response.Extraction-based methods can leverage knowledge more efficiently and be less prone to break complete semantic units,but the extracted responses always lack fluency and naturalness.This dissertation proposes a reference-aware model to combine the advantages of the two kinds of methods mentioned above to improve the performance RG.Specifically,this dissertation designs a decoding switcher and a hybrid decoder.At each decoding step,the decoding switcher decides between reference decoding and generation decoding.Based on the decision made by the decoding switcher,the hybrid decoder either extracts a semantic unit from external knowledge(reference decoding)or generates a token otherwise(generation decoding).Based on generating the response token by token,the reference-aware model also provides an alternative way to learn to extract a semantic unit from external knowledge directly.Experimental results show that the reference-aware model significantly outperforms state-of-the-art methods.