A Generative Model for User Simulation in a Spatial Navigation Domain

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 626–635, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Generative Model for User Simulation in a Spatial Navigation Domain Aciel Eshky1, Ben Allison2, Subramanian Ramamoorthy1, and Mark Steedman1 1School of Informatics, University of Edinburgh, UK 2Actual Analytics Ltd., Edinburgh, UK {a.eshky,s.ramamoorthy,steedman}@ed.ac.uk ballison@actualanalytics.com Abstract We propose the use of a generative model to simulate user behaviour in a novel task- oriented dialog domain, where user goals are spatial routes across artificial land- scapes. We show how to derive an effi- cient feature-based representation of spa- tial goals, admitting exact inference and generalising to new routes. The use of a generative model allows us to capture a range of plausible behaviour given the same underlying goal. We evaluate intrin- sically using held-out probability and per- plexity, and find a substantial reduction in uncertainty brought by our spatial repre- sentation. We evaluate extrinsically in a human judgement task and find that our model’s behaviour does not differ signif- icantly from the behaviour of real users. 1 Introduction Automated dialog management is an area of re- search that has undergone rapid advancement in the last decade. The driving force of this innova- tion has been the rise of the statistical paradigm for monitoring dialog state, reasoning about the effects of possible dialog moves, and planning fu- ture actions (Young et al., 2013). Statistical di- alog management treats conversations as Markov Decision Processes, where dialog moves are as- sociated with a utility, estimated online by inter- acting with a simulated user (Levin et al., 1998; Roy et al., 2000; Singh et al., 2002; Williams and Young, 2007; Henderson and Lemon, 2008). Slot- filling domains have been the subject of most of this research, with the exception of work on trou- bleshooting domains (Williams, 2007) and rela- tional domains (Lison, 2013). Although navigational dialogs have received much attention in studies of human conversational behaviour (Anderson et al., 1991; Thompson et al., 1993; Reitter and Moore, 2007), they have not been the subject of statistical dialog management research, and existing systems addressing naviga- tional domains remain largely hand crafted (Ja- narthanam et al., 2013). Navigational domains present an interesting challenge, due to the dispar- ity between the spatial goals and their grounding as utterances. This disparity renders much of the statistical management literature inapplicable. In this paper, we address this deficiency. We focus on the task of simulating user be- haviour, both because of the important role sim- ulators plays in the induction of dialog managers, and because it provides a self-contained means of developing the domain representations which fa- cilitate dialog reasoning. We show how a genera- tive model of user behaviour can be induced from data, alleviating the manual effort typically in- volved in the development of simulators, and pro- viding an elegant mechanism for reproducing the natural variability observed in human behaviour. 1.1 Spatial Goals of Users Users in task-oriented domains are goal-directed, with a …

Content truncated to 500 words