<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0304">
  <Title>Incremental Parsing with Reference Interaction</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 An Incremental Parsing Architecture
</SectionTitle>
    <Paragraph position="0"> Many current parsers fall into the class of history-based grammars (Black et al., 1992). The independence assumptions of these models make the parsing problem both stochastically and computationally tractable, but represent a simplification and may therefore be a source of error. In a continuous understanding framework, higher-level modules may have additional information that suggests loci for improvement, recognizing either invalid independence assumptions or errors in the underlying probability model.</Paragraph>
    <Paragraph position="1"> We have designed a general incremental parsing architecture (Figure 1) in which the Client, a dynamic programming parser, performs its calculations, the results of which are incrementally passed on via a Mediator to an Advisor with access to higher-level information. This higher-level Advisor sends feedback to the Mediator which has access to the Client's chart, and which then surreptitiously changes and/or adds to the chart in order to make the judgments conform more closely to those of the Advisor. The parser, whose chart has (unbeknownst to it) been changed, then simply calculates chart expansions for the next word, na&amp;quot;ively expanding the currently available (and possibly modified) hypotheses.</Paragraph>
    <Paragraph position="2"> This architecture is general in that neither the Mediator nor the Advisor have been specified; either of these modules can be instantiated in any number of ways within the general framework. The typical dynamic programming component will function in very much the same way that it does in the vanilla algorithm, except that the chart in which partial results are recorded may be modified between time steps. The Client can be any system which uses dynamic programming to efficiently encode independence assumptions, so long as it provides the Mediator with the ability to modify chart probabilities and add chart entries; otherwise the original parser can remain untouched. By having the Mediator perform these modifications rather than the Advisor, we preserve modularity: in this architecture the Advisor need not be aware of the specific implementation of the Client, although depending on the type of advice provided, it may need access to the underlying grammar. The Mediator isolates the Advisor and Client from each other as well as determining how the feedback will be introduced into into the Client's chart.</Paragraph>
    <Paragraph position="3"> Stoness (2004) identifies two broad categories of subversion - our term for the Mediator's surreptitious modification of the Client's chart - as outlined below: + Heuristic Subversion: the Mediator uses the Advisor's feedback as heuristic information, affecting the search sequence but not the probabilities calculated for a given hypothesis; and + Chart Subversion: the Mediator is free to modify the Client's chart as necessary, but does not directly affect the search sequence of the Client (except insofar as this is accomplished by the modifications to the chart).</Paragraph>
    <Paragraph position="4"> The two types of subversion have very different properties. Heuristic subversion will affect the set of analyses which is output by the parser, but each of those analyses will have exactly the same probability score as under the original parser; the effects of the Advisor are essentially limited to determining which hypotheses remain within the beam, or the order in which hypotheses are expanded, depending on whether the underlying parser uses a beam search or an agenda. Chart subversion, on the other hand, will actually change the scores assigned analyses, resulting in a new probability distribution.</Paragraph>
    <Paragraph position="5"> Heuristic subversion is considerably less powerful, but more stable; the effects of chart subversion can be fairly chaotic, especially if care is not taken to avoid feedback loops. Stoness (2004) outlines conditions under which the effects of chart subversion are predictable, becoming broadly equivalent to an incremental version of a post-hoc re-ranking of the Client's output hypotheses.</Paragraph>
    <Paragraph position="6"> Further details on the general architecture, including properties of various modes of feedback integration, a discussion of the relationship between incremental parsing and parse re-ranking, the possibilities of multiple Advisors working in combination, and provisions in the model for asynchronous feedback are available in a University of Rochester Technical Report (Stoness, 2004).</Paragraph>
  </Section>
class="xml-element"></Paper>