Natural Language Generation as Planning Under Uncertainty for Spoken Dialogue Systems

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 683–691, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Natural Language Generation as Planning Under Uncertainty for Spoken Dialogue Systems Verena Rieser School of Informatics University of Edinburgh vrieser@inf.ed.ac.uk Oliver Lemon School of Informatics University of Edinburgh olemon@inf.ed.ac.uk Abstract We present and evaluate a new model for Natural Language Generation (NLG) in Spoken Dialogue Systems, based on statis- tical planning, given noisy feedback from the current generation context (e.g. a user and a surface realiser). We study its use in a standard NLG problem: how to present information (in this case a set of search re- sults) to users, given the complex trade- offs between utterance length, amount of information conveyed, and cognitive load. We set these trade-offs by analysing exist- ing MATCH data. We then train a NLG pol- icy using Reinforcement Learning (RL), which adapts its behaviour to noisy feed- back from the current generation context. This policy is compared to several base- lines derived from previous work in this area. The learned policy significantly out- performs all the prior approaches. 1 Introduction Natural language allows us to achieve the same communicative goal (“what to say”) using many different expressions (“how to say it”). In a Spo- ken Dialogue System (SDS), an abstract commu- nicative goal (CG) can be generated in many dif- ferent ways. For example, the CG to present database results to the user can be realized as a summary (Polifroni and Walker, 2008; Demberg and Moore, 2006), or by comparing items (Walker et al., 2004), or by picking one item and recom- mending it to the user (Young et al., 2007). Previous work has shown that it is useful to adapt the generated output to certain features of the dialogue context, for example user prefer- ences, e.g. (Walker et al., 2004; Demberg and Moore, 2006), user knowledge, e.g. (Janarthanam and Lemon, 2008), or predicted TTS quality, e.g. (Nakatsu and White, 2006). In extending this previous work we treat NLG as a statistical sequential planning problem, anal- ogously to current statistical approaches to Dia- logue Management (DM), e.g. (Singh et al., 2002; Henderson et al., 2008; Rieser and Lemon, 2008a) and “conversation as action under uncertainty” (Paek and Horvitz, 2000). In NLG we have similar trade-offs and unpredictability as in DM, and in some systems the content planning and DM tasks are overlapping. Clearly, very long system utterances with many actions in them are to be avoided, because users may become confused or impatient, but each individual NLG action will convey some (potentially) useful information to the user. There is therefore an optimization prob- lem to be solved. Moreover, the user judgements or next (most likely) action after each NLG action are unpredictable, and the behaviour of the surface realizer may also be variable (see Section 6.2). NLG could therefore fruitfully be approached as a sequential statistical planning task, where there are trade-offs and decisions to make, such …

Content truncated to 500 words