<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1029">
  <Title>Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 225-232, Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Error Handling in the RavenClaw Dialog Management Framework</Title>
  <Section position="7" start_page="231" end_page="231" type="concl">
    <SectionTitle>
5 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> We have described the error handling architecture underlying the RavenClaw dialog management framework. Its design is modular: the error handling strategies as well as the mechanisms for engaging them are decoupled from the actual dialog task specification. This significantly lessens the development effort: system authors focus exclusively on the domain-specific dialog control logic, and the error handling behaviors are generated transparently by the error handling process running in the core dialog engine. Furthermore, we have argued that the distributed nature of the error handling process leads to good scalability properties and facilitates the reuse of policies within and across systems and domains.</Paragraph>
    <Paragraph position="1"> The proposed architecture represents only the first (but an essential step) in our larger research program in error handling. Together with the systems described above, it sets the stage for a number of current and future planned investigations in error detection and recovery. For instance, we have recently conducted an extensive investigation of non-understanding errors and the ten recovery strategies currently available in the RavenClaw framework. The results of that study fall beyond the scope of this paper and are presented separately in (Bohus and Rudnicky, 2005a). In another project supported by this architecture, we have developed a model for updating system beliefs over concept values in light of initial recognition confidence scores and subsequent user responses to system actions. Initially, our confirmation strategies used simple heuristics to update the system's confidence score for a concept in light of the user response to the verification question. We have showed that a machine learning based approach which integrates confidence information with correction detection information can be used to construct significantly more accurate system beliefs (Bohus and Rudnicky, 2005b). Our next efforts will focus on using reinforcement learning to automatically derive the error recovery policies.</Paragraph>
  </Section>
class="xml-element"></Paper>