<?xml version="1.0" standalone="yes"?>
<Paper uid="C00-2112">
  <Title>Making Sense of Reference to the Unfamiliar</Title>
  <Section position="4" start_page="775" end_page="776" type="metho">
    <SectionTitle>
2 A Framework for Reference
Resolution
</SectionTitle>
    <Paragraph position="0"> Our framework for reference resolution has been implemented in the system of language understanding described in (Ramsay, 1999). The starting point tbr reference resolution is the logical tbrm we obtain fl'om parsing. For example, the tbllowing is the logical tbrm we get for the utterance &amp;quot;Mary slept.&amp;quot;</Paragraph>
    <Paragraph position="2"> We use tile inference engine described in (Ramsay and Seville, 2000) to update the discourse model with a new discourse state containing the intbrmation explicitly represented in tile logical tbrm together with any further inferences which are licensed given the existing discourse model. Reference resolution, which involves carrying out a proof that a retbrring expression denotes, is implemented as part of the update step. We anchor a referring expression like ref()~D(named(D, Marg)&amp;card(D, 1))) in tile discourse model by proving the existence of an entity in the model which satisfies the properties specified by the referring expression, in this case aD(na,~ed(D, Mary)~ea,'d(D, 1)) 2.</Paragraph>
    <Paragraph position="3"> 2Strictly speaking, it is a set which is denoted. For readability, our referring expressions conflate tim properties of sets and their members. In this case, the cardinality is a property of the set denoted, but the nmne Mary is a property of its member.</Paragraph>
    <Paragraph position="4"> Given that many referring expressions do not in themselves denote uniquely, however, we need a theory of reference resolution to enable us to obtain the appropriate (i.e., intended) reterent for any referring expression. We incorporate our theory of reference resolution into the actual representation of referring expressions; for example, we label anaphors with the prop-erty &amp;quot;salient&amp;quot; and pronouns (and also referential tenses) with the property &amp;quot;centred&amp;quot;3: &amp;quot;himself&amp;quot; re f ( AX (salient ( X, re f ( AD ( eds (D)))$~m (X))) &amp;quot;she ~, ref (),X (ee..tred(X, reZ (),D(e&amp;(D) ) )~Z(X) ) ) Retbrence resolntion relies on maintaining, as in Centering Theory, a list of tbrward-looking centres for each discourse state (corresponding to an utterance) in the discourse. Furthermore, for the purposes of reference resolution, the discourse states themselves are organized into a discourse tree, which is constructed automatically based on referential cues 4, as described in (Seville, 1999).</Paragraph>
    <Paragraph position="5">  (1) a mani diedj in a park/~.</Paragraph>
    <Paragraph position="6"> (2) hei hadj been sleeping ther%.</Paragraph>
    <Paragraph position="7"> (3) a womanl lovedm him/.</Paragraph>
    <Paragraph position="8"> (4) shez had,~ hated him/.</Paragraph>
    <Paragraph position="9"> (5) he/ hadm hated himself/.</Paragraph>
    <Paragraph position="10"> (6) he~: hadm loved herl.</Paragraph>
    <Paragraph position="11">  The nodes in such a tree correspond to discourse states. Those oll tile right-hand frontier are open, which essentially means that tile entities mentioned in them are available to pronominal reference.</Paragraph>
    <Paragraph position="12"> The process of reference resolution for tile various referring expressions can be briefly described as tbllows. Anaphors, characterised as salient, are resolved to a less oblique argument of the same verb (Pollard and Sag, 1994) within the current discourse state, which is constructed aHere rcf(AD(cds(D)) is a reference to the current discourse state and the properties m and f refer to male and female gender respectively.</Paragraph>
    <Paragraph position="13"> 4The tree illustrated was constructed using pronominal cues. Each discourse state was attached as a daughter of the highest node in the discourse tree to which all pronouns and referential tenses (like had) mentioned in it could be anchored.</Paragraph>
    <Paragraph position="14">  incrementally. We also st;art our sere'oh tbr the referents of prononns and other centred entities in the current disconrse state, which is necessary if we are to resolw; such referring expressions as &amp;quot;her&amp;quot; in &amp;quot;Mary took John with her.&amp;quot; However, referring expressions containing the property centred are prevented front 1)eing dereferenced to salient entities, thus ensuring that the constraint of disjoint reference is met. If we fail to tind the centred entity in the current discourse state, we search the previous open node and, if necessary, fllrther open nodes in the discourse tree, in order to deal with long-distance pronominalisation.</Paragraph>
    <Paragraph position="15"> The dereferencing of other referring expressions like ref(AD(named(D, Mary)gcard(D, 1))) is similar but less constrained in that we consider entities mentioned in all nodes mentioned in the discourse, tree, whether open or closed, in order of recency. This means that, essentially, names and definite descriptions are derefcreated to the most recently mentioned referent which is appropriate. Unlike in the case of pronouns, we also consider Discourse State 0, which doesn't correspond to an utterance but, rather, contains the background knowledge asstoned in the model. This is how we are able to deal with the first mention of a familiar referent like Mary (assmning that the properties kD(na, m(:d( D, Mary)gcard( D , 1)) sumce to distinguish a particular entity in Discourse, State 0 from all the others).</Paragraph>
    <Paragraph position="16"> Our approach extends naturally to cases like %he first snowdrop of Spring&amp;quot; because it; is proof-theoretic and so able to exploit background knowledge in reference resolution. This can be illustrated, in the first instance., by examthing the backgrmmd knowledge which is used in updating the utterance &amp;quot;Mary slept.&amp;quot; The update step for this utterance yields Discourse  State 1, contailfing (amongst others) the tbllowing facts: Discourse State 1 ,s/eep(#134) 0(#134, agent, #94) ends_before(#4(1), #133) aspect(simple, #13a, #134)  We were able to prove named(#94, Mary) and card(#94, 1) and so dereference rcf(,\D(namcd(D, Mary)&amp;card(D, 1))) as the tbllowing were amongst the t~cts contained in Discourse State 0:</Paragraph>
    <Paragraph position="18"> These were generated from the lexical memfing 1)ostulates we stipulated for &amp;quot;Mary&amp;quot;, &amp;quot;woman&amp;quot;,</Paragraph>
    <Paragraph position="20"/>
  </Section>
class="xml-element"></Paper>