MODELLING USER BEHAVIOUR IN THE HIS-POMDP DIALOGUE MANAGER S. Keizer, M. Gaˇsi´c, F. Mairesse, B. Thomson, K. Yu, and S. Young Department of Engineering, University of Cambridge (UK) ABSTRACT
Index Terms— Spoken dialogue systems, POMDPs, user modelling, evaluation 1. INTRODUCTION One of the major challenges in the development of spoken dialogue systems is the problem of uncertainty due to speech recognition (ASR) and language understanding (NLU) errors, and unexpected user behaviour. Most systems rely on choosing the most likely result from ASR and NLU, and in cases where the chosen input turns out to be incorrect, complex repair strategies are required. By modelling dialogue as a POMDP (Partially Observable Markov Decision Process), a framework is provided that enables these uncertainties to be modelled in a well-founded (probabilistic) manner thereby leading to more robust policies [1]. A key component of the POMDP framework is the user act model (UAM) P (au |sm , am ), which represents the probability that the user would generate a user act au , given the last system act am and the system state sm . In this paper, we focus on the design of the UAM in the Hidden Information State (HIS) POMDP system and provide experimental evaluations of its effectiveness. In Section 2, the HIS POMDP system is briefly reviewed, followed in Section 3 by a more detailed discussion of the UAM. In Section 4 evaluation results on both simulated and real data are presented, showing the effectiveness of the UAM. Finally, Section 5 gives some conclusions. This research was partly funded by the UK EPSRC under grant agreement EP/F013930/1 and by the EU FP7 Programme under grant agreement 216594 (CLASSiC project: www.classic-project.org).
The architecture of a POMDP-based dialogue system is shown in Figure 1. The user produces an action au based on his goal su , resulting in a speech signal to be analysed by the system’s speech understanding component. This results in an N-best list of user action hypotheses a1u , . . . , aN u . Instead of just taking the 1-best result to update the dialogue state sm , the entire N-best list with confidence scores is used to compute a probability distribution over all possible dialogue states, the belief state b(sm ). sd Speech Understanding au
au1 N
User a~ m
Belief Estimator
..
In the design of spoken dialogue systems that are robust to speech recognition and interpretation errors, modelling uncertainty is crucial. Recently, Partially Observable Markov Decision Processes (POMDPs) have been shown to provide a well-founded probabilistic framework for developing such systems. This paper reports on the design and evaluation of the user act model (UAM) as part of the Hidden Information State (HIS) POMDP dialogue manager. Within this system, the UAM represents the probability of a user producing a certain dialogue act, given the last system act and the dialogue state. Its design is domain-independent and founded on the notions of adjacency pairs and dialogue act preconditions. Experimental evaluation results on both simulated and real data show that the UAM plays a significant role in improving robustness, but it requires that the N-best lists of user act hypotheses and their confidence scores are of good quality.
ing robustness, but it requires that the N-best lists of user act hypotheses ... Based on the current belief state b, the machine selects an ... user performed his action into account. .... cess rate when using the UAM (the dialogue scores not given.
It is useful to try and build ... person can be useful if trying to build up this picture, although it must be remembered that people ... SEEKING PROFESSIONAL HELP .... All fact sheets can be downloaded free of charge from our website ...
when these theories establish the ideal aims, the technology tradeoffs .... Analyzing Student- Tutor Interaction Logs to Improve Educational Outcomes, (2004).
improving collaboration, particularly by capturing data that can be used to model the nature of the interactions ... a careful analysis of the nature of tabletop datasets and collaboration. Keywords: .... There are also software based solutions that
Cytological studies in interspecific hybrid derivatives of cotton viz., IS-244/4/1 and IS-181/7/1 obtained in BC1F8 generation of trispecies cross ... Chromosome association of 5.19 I + 8.33 II +1.14III + 1.09IV and 6.0 I+ 7.7 II +0.7III + 1.25IV was
Oct 27, 2004 - Mall, University of Hawaii at Manoa, Honolulu, HI 96822, U.S.A.. (email: [email protected]).H. L. Pratt is at the Mote Marine. Laboratory, 24244 ..... 211â237. Seattle, Washington: U.S. Department of Commerce, National. Oceanic and
tive entertainment, and in computer video games in particular. ... ment (which amount to directing a virtual avatar to take various actions or speak certain lines of ...
tive entertainment, and in computer video games in particular. Computer ... This paper contains some text from a current demonstration submission to AIIDE'08.
organizational commitment, supervisory support .... autonomy, trust, cohesiveness, support, recognition, ...... Manchester Institute of Science and Technology,.
May 14, 2003 - the RuO2 volume fraction dependence of the piezoresistance data fit well .... of TFRs and its microscopic origin by re-analysing earlier data on ...
Journal of Physics: Conference Series 94 (2008) 012015 ... Cu-Kα1 and Kα2 radiation, was fitted to each peak in the range from 2θ = 12- 75°. A least means.
E-mail: [email protected] ... which are in solution at high temperatures segregate to grain boundaries during slow cooling. .... indicates that it contains less substitutional niobium in the bulk ceramic compared to the ...
1999 Elsevier Science B.V. All rights reserved. Keywords: Aggression .... the data analysis: Spearman rank correlation co- efficient with exact P values based on ...
OOOl-6918/93/$06.00 0 1993 - Elsevier Science Publishers B.V. All rights reserved ..... tion (over the field of body-icons) of the degree of compliance with the different constraints that ... 223. X. Fig. 2. Cortical map of a 2 dof arm after training
straints on how it operates, or may influence the way that it chooses to achieve its ... We claim the same is true of situations. 1049 ... to be true. Figure 2 depicts the derivation rules for cap- turing when a particular situation becomes active an
robust and coherent maps of the different motor spaces) with relaxation dynamics (for run-time incorporation of task constraints) and non-linear integration (for a ...
models are also provided with isotropic hardening. In this case plastic flow causes the yield surface. to change size uniformly with respect to all stress directions.
the clarity and brevity of the presentation. A special feature of this book is in exploring thermomechanical consistency of all presented constitutive models in a ...