Optimising Incremental Dialogue Decisions Using Information Density for Interactive Systems

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 82–93, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Optimising Incremental Dialogue Decisions Using Information Density for Interactive Systems Nina Dethlefs, Helen Hastie, Verena Rieser and Oliver Lemon School of Mathematical and Computer Sciences Heriot-Watt University, Edinburgh, Scotland n.s.dethlefs | h.hastie | v.t.rieser | o.lemon@hw.ac.uk Abstract Incremental processing allows system design- ers to address several discourse phenomena that have previously been somewhat neglected in interactive systems, such as backchannels or barge-ins, but that can enhance the re- sponsiveness and naturalness of systems. Un- fortunately, prior work has focused largely on deterministic incremental decision mak- ing, rendering system behaviour less flexible and adaptive than is desirable. We present a novel approach to incremental decision mak- ing that is based on Hierarchical Reinforce- ment Learning to achieve an interactive op- timisation of Information Presentation (IP) strategies, allowing the system to generate and comprehend backchannels and barge-ins, by employing the recent psycholinguistic hy- pothesis of information density (ID) (Jaeger, 2010). Results in terms of average rewards and a human rating study show that our learnt strategy outperforms several baselines that are not sensitive to ID by more than 23%. 1 Introduction Recent work on incremental systems has shown that adapting a system’s turn-taking behaviour to be more human-like can improve the user’s experience significantly, based on incremental models of auto- matic speech recognition (ASR) (Baumann et al., 2011), dialogue management (Buss et al., 2010), and speech generation (Skantze and Hjalmarsson, 2010). All of these approaches are based on the same gen- eral abstract architecture of incremental processing (Schlangen and Skantze, 2011). While this archi- tecture offers inherently incremental mechanisms to update and revise input hypotheses, it is affected by a number of drawbacks, shared by determinis- tic models of decision making in general: they rely on hand-crafted rules which can be time-consuming and expensive to produce, they do not provide a mechanism to deal with uncertainty introduced by varying user behaviour, they are unable to gener- alise and adapt flexibly to unseen situations, and they do not use automatic optimisation. Statisti- cal approaches to incremental processing that ad- dress some of these problems have been suggested by Raux and Eskenazi (2009), who use a cost matrix and decision theoretic principles to optimise turn- taking in a dialogue system under the constraint that users prefer no gaps and no overlap at turn bound- aries. Also, DeVault et al. (2009) use maximum en- tropy classification to support responsive overlap in an incremental system by predicting the completions of user utterances. Selfridge et al. (2011) use logis- tic regression models to predict the stability and ac- curacy of incremental speech recognition results to enhance performance without causing delay. For re- lated work on (deterministic) incremental language generation, please see (Kilger and Finkler, 1995; Purver and Otsuka, 2003). Recent years have seen a number of data-driven approaches to interactive systems that automatically adapt their decisions …

Content truncated to 500 words