<?xml version="1.0" standalone="yes"?>
<Paper uid="P92-1004">
  <Title>THE REPRESENTATION OF MULTIMODAL USER INTERFACE DIALOGUES USING DISCOURSE PEGS</Title>
  <Section position="6" start_page="28" end_page="29" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> The three-tiered discourse representation was used to model dialogue interaction from one agent's point of view. The discourse pegs level is independent of both the surface forms that occur and the immediate condition of the supporting belief system. In the implemented UI systems the discourse model provided a necessary buffer between the Cyc KB undergoing revision and the ongoing dialogue. However, most of the relevant considerations apply to other HCI dialogues, to human-human dialogues, and to NL discourse processing in general. I summarize the advantages of the pegs model under the original three headings and close with suggestions for further work.</Paragraph>
    <Paragraph position="1"> (A ) Cognitive considerations: The belief system (KB) can serve dialogue processes as a source of information about the reference world without being itself modified as a necessary side effect of discourse interpretation. This means that understanding is not equated with believing, i.e., mismatch between pegs and KB objects is tolerated.</Paragraph>
    <Paragraph position="2"> Separate processes are allowed to update the KB in the background during discourse processing as the represented world changes and afterward, 'belief acquisition' can take care of assimilating pegs into the KB where appropriate.</Paragraph>
    <Paragraph position="3"> The separation of tiers allows for differential rates of information decay. The linguistic tier fades from availability rapidly and as a function of time, discourse tier decay is conditioned by attentional focus, and the KB represents a static belief structure in which forgetting, if represented at all, is not affected by discourse processing.</Paragraph>
    <Paragraph position="4"> Interpretation can be accomplished incrementally.</Paragraph>
    <Paragraph position="5"> The meaning of an NP is not defined as a KB object it corresponds to but as the peg that it mentions in the discourse model, and that peg is always a partial representation of the speaker's intended referent. How partial it is can vary over time and it can be of use for  sponsoring dependent NPs, generating questions, etc., even in its partial state. Indeed, feedback from such use is what helps to further specify the peg.</Paragraph>
    <Paragraph position="6"> (B ) Linguistic phenomena: In English, all NPs have the potential of being context-dependent. The separation of tiers allows for the distinction between true anaphoric dependence and incidental coreference, encoded as the co-anchoring of multiple LOs to a single peg without sponsorship.</Paragraph>
    <Paragraph position="7"> Partial and total anaphors are explicitly represented, with linguistic sponsoring distinguished from discourse sponsoring, and these relationships are stored as annotated links in the permanent discourse representation so that internal NL and non-NL procedures may query the discourse structure for information on coreference, KB property values, justifications for later links, etc.</Paragraph>
    <Paragraph position="8"> The distinction between discourse and linguistic sponsoring allows language-specific syntactic and semantic constraints to be upheld at the LO level and overridden by pragmatic and discourse considerations at the discourse pegs level, thereby providing a mechanism which addresses well-known violations of linguistic constraints on coreference without relaxing the constraints themselves.</Paragraph>
    <Paragraph position="9"> Input and output are distinguished at the linguistic tier but merged at the discourse model tier. The user can make anaphoric reference through any channel to pegs introduced by the backend system through any channel. Yet it remains part of the discourse history record in the linguistic tier, who made which assertions about which pegs. In the HCI dialogue environment this means that NL and non-NL modalities are equally acceptable as surface forms for input and output utterances, i.e., voice input could be added without extension to the current system as long as the speech recognizer output forms that could be used to generate LOs.</Paragraph>
    <Paragraph position="10"> ( C) Evidence from a trial implementation: In knowledge-based UIs, the strict separation of tiers means that the KB can be incomplete or incorrect throughout the discourse, it can remain unaffected by discourse processing, and it can be updated by other knowledge acquisition procedures independently of simultaneous discourse processing.</Paragraph>
    <Paragraph position="11"> Nevertheless, it is possible and may be computationally efficient to implement the discourse model as a specialized, non-static (and potentially redundant) region of the KB so that KB reasoning mechanisms can be applied to the hypothetical state of affairs depicted by pegs in the discourse model.</Paragraph>
    <Paragraph position="12"> The guise of an individual has just those properties assumed by the current discourse. Using pegs as dynamically defined guises in effect suppresses non-salient properties of the accessed KB unit. Thus Grosz's requirement that the discourse representation encode relations in focus as well as entities in focus is supported at the pegs level.</Paragraph>
    <Paragraph position="13"> Moreover, the three-tiered design can represent conflict between interpreted discourse information and the agent's static beliefs because KB values can be overridden in the discourse by ascription of contrary properties to corresponding pegs. A related benefit is that the external dialogue participant is allowed to introduce new pegs and new information into the discourse and this does not require creation of a new KB object during discourse interpretation.</Paragraph>
    <Paragraph position="14"> Because pegs are used to accumulate tentative properties on (actual or hypothetical) individuals without editing the KB either permanently or only for the duration of the discourse, belief acquisition can be postponed until a sufficiently complete understanding has been achieved, so the discourse model can serve as an agenda for later KB updating. Meanwhile, partial and incorrect discourse representations are useful and non-monotonic repair operations make it easy to correct interpretation errors by changing links between LO and peg or between peg and KB unit without disturbing other links.</Paragraph>
    <Paragraph position="15"> Some pegs are not associated with the linguistic tier at all. Graphical events in the physical environment that make an object salient can inject a peg directly into the discourse model. However, only pegs introduced via the linguistic channel can sponsor linguistic anaphora, e.g., &amp;quot;What is it&amp;quot; requires the presence of an LO, but &amp;quot;What is that&amp;quot; can be sponsored directly by the peg for an icon that just appeared on the screen.</Paragraph>
    <Section position="1" start_page="29" end_page="29" type="sub_section">
      <SectionTitle>
Further Research
</SectionTitle>
      <Paragraph position="0"> Dependents can sponsor other dependents, and in general, there is complex interaction between sequences of NPs in a discourse. For example, in the sentence Delete the buttons if one of them is missing its label. its label is partially dependent on one, and/t is totally dependent on 9nC/ which is partially dependent on them which is totally dependent on the buttons which is presumably a total anaphoric reference to a discourse peg for some set of buttons currently in focus. The present algorithm attempts pseudo-parallel processing of LOs, taking repeated passes through the new utterance, left to right by NP type, (proper nouns, definite NPs,..,reflexives). One-anaphors modified by partitive PPs are exceptional in that they are processed after the pronoun or definite NP (the object of the preposition) to their right. Further work is needed to describe the ways that various NP types interact as this was a technique for coping with the absence of a theory of the possible relationships between sequences of partial and total anaphoric NPs.</Paragraph>
      <Paragraph position="1"> LOs for events are created by the semantic processing module and so sequences such as: You deleted that unit. I didn't want to do that.</Paragraph>
      <Paragraph position="2"> could in theory be handled analogously with other partial and total anaphors. However, they are not of use in the current application UIs and so their theory and implementation have remained undeveloped here.</Paragraph>
      <Paragraph position="3"> Ambiguous mouse clicks of the sort explored in XTRA and CUBRICON plus the ability of the user to introduce new pegs for regions of the screen, or for events of moving a pane or icon across the screen, or encircling a set of existing icons to place their pegs in attentional focus should all be attempted using the pegs discourse model as a source of target interpretations of mouse clicks and as a place to encode novel, user-defined screen objects.</Paragraph>
      <Paragraph position="4"> Finally, with this or other representations of dialogue, a variety of UI metaphors should be explored. The UI can be viewed as a single autonomous agent or as merely the clearing house for communication between the user and a collection of agents, the operating system, the graphical interface, the NL system, or any of the knowledge sources, such as those on the HITS blackboard, which could conceivably want to engage the user in a dialogue.</Paragraph>
      <Paragraph position="5"> The three-tiered discourse design is also used in the knowledge based NL system at MCC (Barnett, et al., 1990), and is being explored as one descriptive device for dialogue in voice-to-voice machine * translation at ATR.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>