<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1306">
  <Title>Multidimensional Dialogue Management</Title>
  <Section position="7" start_page="42" end_page="43" type="concl">
    <SectionTitle>
5 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> We have presented a dialogue management approach supporting the generation of multifunc- null alogue theory involving a multidimensional dialogue act taxonomy and an information state on which the dialogue acts operate. Several dialogue acts from different dimensions are generated by dialogue act agents associated with these dimensions, and can thus be combined into multifunctional system utterances.</Paragraph>
    <Paragraph position="1"> A first implementation of a dialogue manager following this multi-agent approach has been integrated into an interactive QA system and supports a limited range of dialogue acts from the DIT taxonomy, both for interpreting user utterances and generating system utterances. The system is able to attend to different aspects of the communication simultaneously, involving reactive social behaviour, answering domain questions and giving feedback about utterance interpretation and the question-answering process.</Paragraph>
    <Paragraph position="2"> Future development will involve extending the range of dialogue acts to be covered by the dialogue manager, for a part following from the definition of an extended system functionality, and consequently, extending the set of dialogue act agents. This also has consequences for the Evaluation Agent: the process of combination and selection will be more complex if more dialogue act types can be expected and if the dialogue acts have a semantic content that is more than just a collection of QA-answers.</Paragraph>
    <Paragraph position="3"> In terms of system functionality we aim at sup- null port for generating articulate feedback, i.e., feed-back acts that are not merely signalling processing success or failure, but (in case of negative feedback) also contain a further specification of the processing problem at hand. For example, the system may have encountered problems in processing certain parts of a user utterance, or in resolving an anaphor; then it should be able to ask the user a specific question in order to obtain the information required to solve the processing problem (see the example in Section 4.1.2). The articulate feed-back acts may also involve dealing with problems in the question answering process, where the system should be able to give specific instructions to the user to reformulate his question or give additional information about his information need.</Paragraph>
    <Paragraph position="4"> In addition to supporting generation of articulate feedback acts, we also aim at dialogues between user and system that are more coherent and natural, i.e., the system should be more aware of the conversational structure, and display more refined social behaviour. Not only should it generate simple reactions to greetings, apologies, and goodbyes; it should also be able to generate initiative social acts, for example, apologies after several cases of negative auto-feedback.</Paragraph>
    <Paragraph position="5"> The extended set of dialogue acts will also lead to an extended context model. Related to the context model and updating mechanism is on-going work on belief dynamics and grounding in DIT (Morante and Bunt, 2005). The defined mechanisms for the creation, strengthening, adoption, and cancelling of beliefs and goals in the context model are currently being implemented in a demonstrator tool and will also be integrated in the information state update mechanism of the PARADIME dialogue manager.</Paragraph>
  </Section>
class="xml-element"></Paper>