Proactive Mixed-type Dialogue Systems

A dialogue system is a human-computer interaction system.Dialogue systems can generally be divided into chit-chat and task-oriented dialogue.Chit-chat is open-domain dialogue and generally has no clear dialogue goal,while task-oriented dialogue aims at specific domains and has clear dialogue goals to achieve.In recent years,the academic community has made significant progress in research on different types of dialogues.The industry has also released many dialogue system-related products.However,previous researchers always studied different dialogue types separately and have not yet considered how to naturally integrate them(called "mixed-type dialogue").Moreover,human-machine dialogue in many real-world applications generally involves multiple dialogue types.For example,a user wants to travel to the Forbidden City,but he does not know it well.In this case,the dialogue systems need to help the user know the Forbidden City through chit-chat,and then book tickets through task-oriented dialogue.Therefore,mixed-type dialogue is closer to real-world scenarios than the previous research.According to the different roles of guiding dialogues,dialogue systems can be divided into dialogue systems with bot proactive or user proactive.The dialogue systems with the user proactive are dependent on the proactivity of the user.If the user is no longer proactive,or he does not know how to guide,the dialogues may not continue.Therefore,to better meet user needs,the dialogue system needs to proactivity and naturally guide dialogues.Based on the consideration of the bot’s proactivity and naturally integrating of common dialogue types,this thesis studies mixed-type dialogue with the bot proactive(called "proactive mixed-type dialogue").Specifically,this thesis first proposes two new proactive mixed-type dialogue tasks for open-domain conversational recommendation or task-oriented dialogue,and studies them in an end-to-end manner.Then,in order to improve the controllability of mixed-type dialogue,the above tasks are divided into a dialogue goal planning module and a response generation module to introduce more supervised signals and promote the generation of mixed-type dialogue.Finally,this thesis studies goal-guided proactive multilingual mixed-type dialogue tasks to make them widely applicable in different countries.The research of this thesis mainly includes the following four aspects:1.Proactive Mixed-type Dialogue for Conversational Recommendation.In view of the problem that only focused on a single dialogue type in the previous research on conversational recommendation,this thesis proposes a recommendation-oriented proactive mixed-type dialogue task based on the guidance from non-recommendation dialogues to recommendation dialogues and constructs a corresponding dataset.A dialogue generation model with a novel mixed-goal-driven dialog mechanism is proposed to address this task.The experiment demonstrates the effectiveness of this model.2.Proactive Mixed-type Dialogue for Task Completion.In view of the problem,users are supposed to have explicit goals in the previous research of task-oriented dialogue,which is inconsistent with the actual situation.To address this issue,this thesis proposes a proactive mixed-type dialogue task for task completion from the perspective of helping users clarify their goals and constructs a corresponding dataset.To address this task,a dialogue model with a novel prompt-based continual learning mechanism is proposed.The experiment demonstrates the effectiveness of this model.3.Hierarchical Graph-Grounded Goal Planning for Proactive Mixed-type Dialogue.To address the issues of poor controllability in mixed-type dialogue generation,this thesis proposes a goal-planning model for proactive mixed-type dialogue based on the hierarchical graph from the perspective of improving dialogue management.The experiment shows that this model can improve the controllability of proactive mixed-type dialogue systems,thereby enhancing the dialog proactivity and naturalness,and goal completion rate in proactive mixed-type dialogue systems.4.Goal-guided Response Generation for Multilingual Proactive Mixed-type Dialogue.To address the issues of monolingual settings in mixed-type dialogue generation,this thesis is the first to study goal-guided multilingual proactive mixed-type dialogue and proposes a proactive multilingual mixed-type dialogue generation model based on K-nearest neighbor search,which uniformly supports monolingual,multilingual,and cross-lingual response generation.The experiment demonstrates the effectiveness of this model.In general,this thesis studies how to naturally integrate common dialogue types to facilitate the research of proactive mixed-type dialogue.Proactive mixed-type dialogue can improve the performance of dialogue and has important scientific research and industrial application value.We hope that the research results of this thesis can be further expanded to more dialogue scenes and tasks,to promote the further development of dialogue systems.