Generative Goal-Driven User Simulation for Dialog Management

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 71–81, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Generative Goal-Driven User Simulation for Dialog Management Aciel Eshky ILCC School of Informatics University of Edinburgh a.eshky@sms.ed.ac.uk Ben Allison ILCC School of Informatics University of Edinburgh {ballison, steedman}@inf.ed.ac.uk Mark Steedman ILCC School of Informatics University of Edinburgh Abstract User simulation is frequently used to train statistical dialog managers for task-oriented domains. At present, goal-driven simula- tors (those that have a persistent notion of what they wish to achieve in the dialog) re- quire some task-specific engineering, making them impossible to evaluate intrinsically. In- stead, they have been evaluated extrinsically by means of the dialog managers they are in- tended to train, leading to circularity of argu- ment. In this paper, we propose the first fully generative goal-driven simulator that is fully induced from data, without hand-crafting or goal annotation. Our goals are latent, and take the form of topics in a topic model, clustering together semantically equivalent and phoneti- cally confusable strings, implicitly modelling synonymy and speech recognition noise. We evaluate on two standard dialog resources, the Communicator and Let’s Go datasets, and demonstrate that our model has substantially better fit to held out data than competing ap- proaches. We also show that features derived from our model allow significantly greater im- provement over a baseline at distinguishing real from randomly permuted dialogs. 1 Introduction Automatically simulating user behaviour in human- machine dialogs has become vital for training sta- tistical dialog managers in task-oriented domains. These managers are often trained with some vari- ant of reinforcement learning (Sutton and Barto, 1998), where optimal behaviour is sought or learnt through the exploration of the space of possible di- alogs. Although learning by interacting with human subjects is a possibility (Ga˘si´c et al., 2011), it has been argued that user simulation avoids the expen- sive, labour intensive, and error-prone experience of exposing real humans to fledgling dialog systems (Eckert et al., 1997). Training effective dialog managers should benefit from exposure to properties exhibited by real users. Table 1 shows an example dialog in a domain such as we consider, where the objective is to simulate at the semantic level. In such task oriented domains, the user has a goal (in this case, to book a flight from New York to Osaka), and the machine is tasked with fulfilling it. Notice that the user is consistent with this goal throughout the dialog, in that they do not provide contradictory information (although an ASR error is present), but that every mention of their destination city uses a different string. This moti- vates our first desideratum: that simulation be con- sistent over the course of a dialog. Furthermore, one can imagine users not always responding identically in identical situations: we thus additionally require variability. In this paper we demonstrate a fully gen- erative, latent variable probability model exhibiting both of these properties. Thus far, …

Content truncated to 500 words