Conversational AI 23-TASK-ORIENTED NEURAL DIALOGUE SYSTEMS
Initially, end-to-end dialogue learning was applied to open-domain conversational (or chit-chat) systems with promising results. More recently, it has also been applied to task-oriented dialogue systems. However, task-oriented systems present additional challenges. Assuming that the system is able to correctly identify the task that the user wants the system to help solve, it is not just a matter of finding an appropriate response. The system may have to generate a series of questions to elaborate on or clarify the user’s request, issue one or more queries to an external source such as a database to obtain the required information, present the options retrieved back to the user, and then complete the transaction. Moreover, whereas developers of dialogue systems are able to draw on a range of existing corpora of open-domain conversations to train their systems, there are fewer datasets for task-oriented applications and often they are proprietary and not publicly available. Even where corpora of task-oriented dialogues are available, it is necessary to distinguish between generic features that can be learned and those features that are only applicable in a particular domain.