<?xml version="1.0" standalone="yes"?>
<Paper uid="P89-1032">
  <Title>A COMPUTATIONAL MECHANISM FOR PRONOMINAL REFERENCE</Title>
  <Section position="4" start_page="263" end_page="266" type="intro">
    <SectionTitle>
3. THE ALGORITHM
</SectionTitle>
    <Paragraph position="0"> Before giving the details of the algorithm, we will sketch its general structure. The algorithm applies to a completed parse tree and traverses it in a left-toright, depth-first manner. The algorithm uses the notion of minimal domain introduced in the preceding section: the S node or NP node (when minimality has been induced by the presence of a possessive) that most immediately dominates the node being processed, and the related notions of &amp;quot;internal&amp;quot; and &amp;quot;external&amp;quot; nodes. Internal nodes are dominated by the current minimal domain node; external nodes c-command the current minimal domain node. Essentially, the algorithm passes each node all the nodes that c-command it, subdivided into two sets, those that are internal to the current minimal domain and those that are external. As each node is processed, a subroutine is called that dispatches on the category of the node and performs any actions that are appropriate. It is this subroutine that implements the pronominal reference mechanism proper.</Paragraph>
    <Paragraph position="1"> Given this overview, we can now turn to the data structures that are used by the algorithm, as well as to the details of the algorithm. Each node in a parse tree is a Common LISP structure; two of its slots are used for establishing pronominal reference: :possible-antecedentsma list of all the nodes that can be co-referent or overlapping in reference with it.</Paragraph>
    <Paragraph position="2"> :lmpossible-antecedentsBa list of all the nodes that are disjoint in reference with it.</Paragraph>
    <Paragraph position="3"> The algorithm also uses two global variablesB*table-of-proforms* and *table-of-antecedents*rain a &amp;quot;blackboard&amp;quot; fashion. The algorithm uses two major procedures. The first, pass-down-c-commanding-nodes, is responsible for actually traversing each node in the tree. The actual algorithm it uses is shown in Figure 6-1 in a LISP-type notation. Its functionality can be stated as follows. Whenever it encounters a new node, it first processes that node by calling the procedure update-node, which will be described shortly. It next determines whether the node being processed counts as a minimal domain for its children. When the node is a finite S node, it does count as a minimal domain, for all its children. Hence, only nodes that it dominates can be internal nodes for its children; all other nodes are now treated as external by its children. When the node is an NP, there are two possibilities. If there is no possessive NP, the NP does not count as a minimal domain, hence, the external nodes remain as before and the nodes it dominates are added to the set of internal nodes.</Paragraph>
    <Paragraph position="4"> However, when the NP does contain a possessive, it does count as a minimal domain, for all the nodes that it dominates, except the possessive itself. 3 Finally, if the node is of any other category, it is not a minimal domain, so the external nodes remain as before and the internal nodes are augmented by the constituents it dominates. 4 In all cases, pass-down-c-commanding-nodes calls itself recursively on the children of the node being processed, with the appropriate lists of internal and external nodes as arguments.</Paragraph>
    <Paragraph position="5"> update-node, in turn, processes the node passed ~rhe reason for this exception will be explained in Section 6. 4Non-finite clauses also need special treatment. However, consideration of this case requires discussion of whether non-finite clauses are Ss or VPs, which is beyond the scope of this paper.  to it, on the basis of the nodes internal and external to the current minimal domain. In particular, update-node performs the correct pronominal assignment. The algorithm used by update-node is shown in Figure 6-2 in a LISP-type notation. We also discuss each clause separately.</Paragraph>
    <Paragraph position="6"> Clause \[I\] implements condition 1 (non-pronominal NPs). Since there are no minimality conditions on disjoint reference for non-pronominal NPs, all NP nodes c-commanding a non-pronominal NP are added to its :impossible-antecedents slot, whether they are internal (\[I.A\]) or external to the current minimal domain (\[I.B\]). This handles sentences such as those in (9) and (12). While it might seem odd to specify that a non-pronominal NP has no antecedents, this information is useful in handling cases of backwards pronominalization, as in (18).</Paragraph>
    <Paragraph position="7"> (18) \[His\] mother loves \[John\], Clause \[I.C\] handles backwards pronominalization by making use of information in degtable-of-proforms*, a table of all the pronouns encountered so far in the course of the tree walk. s After update-node has added all c-commanding NP nodes to the :impossible-antecedents slot of a non-pronominal NP, it then searches *table-of-proforms* for any pronouns that are not on its :impossible-antecedents list; whenever it finds one, it adds the current non-pronominal NP to the pronoun's :possible-antecedents list. The last thing update-node does in processing a non-pronominal NP is to add it to *table-of-antecedents* (\[I.D\]), whose use will be explained shortly.</Paragraph>
    <Paragraph position="8"> Clause \[11\] implements condition 2 (bound anaphors). Since bound anaphors are short-distance anaphors, all and only the c-commanding NPs internal to the current minimal domain are added to the :possible-antecedents slot of a bound anaphor.</Paragraph>
    <Paragraph position="9"> Clause \[111\] implements condition 3 (personal pronouns). Since personal pronouns are long-distance anaphors, clause \[111\] performs a number of operations. First, all the c-commanding NPs internal to the current minimal domain are added to the :impossible-antecedents slot of a personal pronoun (\[Ill.A\]), disallowing them as antecedents. Next, all the c-commanding NPs external to the current minimal domain are added to the :possible-antecedents slot of a personal pronoun (\[Ill.B\]), indicating that they are potential antecedents. Clause \[Ill.C\] handles sentences like (19).</Paragraph>
    <Paragraph position="10"> (19) \[John's\] mother loves \[him\].</Paragraph>
    <Paragraph position="11"> in which a non-pronominal NP that does not c-command a personal pronoun serves as its antecedent. As was noted above, each non-pronominal NP is added to the *table-of-antecedents* by clause \[I.D\]. When update-node has added all the ap~'his lalok) is filled in by Clause \[Ill.D\].</Paragraph>
    <Paragraph position="12"> propriate c-commanding nodes to the :impossible-antecedents slot of a personal pronoun, it then adds any NPs on *table-of-antecedents* that are not already on the pronoun's :impossible-antecedents slot to its :possible-antecedents slot. Finally, when update-node is finished processing a pronominal NP node, it adds it to *table-of-proforms (Jill.D\]), for use in backwards pronominalization.</Paragraph>
    <Paragraph position="13"> Note that, because our algorithm both establishes minimal domains and assigns possible and impossible antecedents during the course of the tree traversal, it can be single pass, in contrast to Chomsky's procedure, which assigned impossible antecedents in one traversal and checked for minimality during a second.</Paragraph>
    <Paragraph position="14"> Since update-node is a general mechanism for adding or modifying information to a node on the basis of c-commanding constituents it is fairly straightforward to extend to handle other phenomena that involve c-command by modifying its top level CASE statement to dispatch on other categories. In fact, we have extended it in this manner to handle examples of &amp;quot;N anaphora&amp;quot;; i.e. cases where the head noun of a Noun Phrase is either &amp;quot;one&amp;quot; (which has been argued in Baker (1978) to be an anaphor for Ns, i.e. a noun and its complements, but not for full Noun Phrases) or phonologically null (0), which seems to have the same possibilities for antecedents.</Paragraph>
    <Paragraph position="15"> (20) Give me a list of ships which are in the gulf of Alaska that have casualty reports dated earlier than Esteem's oldest one.</Paragraph>
    <Section position="1" start_page="264" end_page="266" type="sub_section">
      <SectionTitle>
Anaphora
</SectionTitle>
      <Paragraph position="0"> The addition to the algorithm that deals with this phenomenon is presented in Figure 3-1. This clause is considerably simpler that those that handle disjoint reference and co-reference phenomena for personal pronouns: only external nodes are involved and only forward antecedence is possible\] This c_lause finds all the Noun Phrases that c-command an N pro-form and that are external to the current minimal domain. This excludes the possessive in a Noun Phrase such as &amp;quot;Esteem's oldest one&amp;quot; or &amp;quot;Wichita's 9&amp;quot; from serving  m as the antecedent to its pro-N. External NPs that meet this criterion are filtered, since not all NPs can be antecedents of an N anaphor. For example, proper nouns cannot serve as such antecedents.</Paragraph>
      <Paragraph position="1"> Each NP that meets these criteria has its N-BAR added to the :possible-antecedents slot of the N-BAR node being processed.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>