Optimising Information Presentation for Spoken Dialogue Systems

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1009–1018, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Optimising Information Presentation for Spoken Dialogue Systems Verena Rieser University of Edinburgh Edinburgh, United Kingdom verena.rieser@ed.ac.uk Oliver Lemon Heriot-Watt University Edinburgh, United Kingdom o.lemon@hw.ac.uk Xingkun Liu Heriot-Watt University Edinburgh, United Kingdom x.liu@hw.ac.uk Abstract We present a novel approach to Informa- tion Presentation (IP) in Spoken Dialogue Systems (SDS) using a data-driven statis- tical optimisation framework for content planning and attribute selection. First we collect data in a Wizard-of-Oz (WoZ) ex- periment and use it to build a supervised model of human behaviour. This forms a baseline for measuring the performance of optimised policies, developed from this data using Reinforcement Learning (RL) methods. We show that the optimised poli- cies significantly outperform the baselines in a variety of generation scenarios: while the supervised model is able to attain up to 87.6% of the possible reward on this task, the RL policies are significantly better in 5 out of 6 scenarios, gaining up to 91.5% of the total possible reward. The RL policies perform especially well in more complex scenarios. We are also the first to show that adding predictive “lower level” fea- tures (e.g. from the NLG realiser) is im- portant for optimising IP strategies accord- ing to user preferences. This provides new insights into the nature of the IP problem for SDS. 1 Introduction Work on evaluating SDS suggests that the Infor- mation Presentation (IP) phase is the primary con- tributor to dialogue duration (Walker et al., 2001), and as such, is a central aspect of SDS design. During this phase the system returns a set of items (“hits”) from a database, which match the user’s current search constraints. An inherent problem in this task is the trade-off between presenting “enough” information to the user (for example helping them to feel confident that they have a good overview of the search results) versus keep- ing the utterances short and understandable. In the following we show that IP for SDS can be treated as a data-driven joint optimisation prob- lem, and that this outperforms a supervised model of human ‘wizard’ behaviour on a particular IP task (presenting sets of search results to a user). A similar approach has been applied to the problem of Referring Expression Generation in di- alogue (Janarthanam and Lemon, 2010). 1.1 Previous work on Information Presentation in SDS Broadly speaking, IP for SDS can be divided into two main steps: 1) IP strategy selection and 2) Content or Attribute Selection. Prior work has presented a variety of IP strategies for structur- ing information (see examples in Table 1). For ex- ample, the SUMMARY strategy is used to guide the user’s “focus of attention”. It draws the user’s at- tention to relevant attributes by grouping the cur- rent results from the database into clusters, e.g. (Polifroni and Walker, 2008; Demberg and Moore, 2006). Other studies investigate a COMPARE strat- egy, e.g. (Walker et al., …

Content truncated to 500 words