<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-2005">
  <Title>Squibs and Discussions Parsing and Empty Nodes</Title>
  <Section position="7" start_page="297" end_page="298" type="concl">
    <SectionTitle>
6. Conclusions
</SectionTitle>
    <Paragraph position="0"> It has not been our purpose here to solve the problem of parsing for GB, but only to provide a mechanism for ensuring that empty nodes do not cause nontermination of parsing in an important class of cases. We have made only very general remarks on the architecture of a parsing system that would incorporate these ideas, largely because we believe that the details of such a design would depend heavily on the mechanism that was chosen for managing constraints. Efficient implementation would depend on a good resolution of a number of interacting trade-offs, and there are several of these within our scheme that need to be explored. In particular, the components of an ELI could be more or less narrowly specified for the roles they are to fill. If the nodes are highly specialized, there will be greater ambiguity in the lexicon and consequently greater nondeterminism in the parser. On the other hand, many of these search paths will presumably be curtailed earlier than they would have been with less specialized nodes.</Paragraph>
    <Paragraph position="1"> A major determinant of system performance will clearly be the manner in which constraints are enforced. It is possible to distinguish a class of constraints that arise in the course of parsing but which cannot, in general, be discharged there, and should therefore be treated as part of the result that the parser delivers. Notable among these are contraindexing constraints from the Binding theory.</Paragraph>
    <Paragraph position="2"> Ensuring that each node in an ELI fills the role for which it was intended could be resolved through the general constraint mechanism. However, more specialized mechanisms could sometimes be useful. Suppose, for example, that the lexical entry for a noun contained a node specifically intended to receive Case. If these were the only nodes whose Case attribute was unspecified, all others having an explicit zero value, the required mechanism could consist simply in having all rules assign a value to this feature, that value being zero except for rules that assign a substantive Case.</Paragraph>
    <Paragraph position="3"> A somewhat different problem consists of verifying that nodes from a given ELI appear in a certain structural configuration. Assigning each node a unique identifier allows this problem to be solved straightforwardly by the general constraint mechanism. null It might be advantageous for the ELI to encode very specific information about a lexical item and the empty nodes that it sponsors. For example, the ELI for a WH  Mark Johnson and Martin Kay Parsing and Empty Nodes item might specify that the traces it sponsors are coindexed with the WH item itself. Assuming that indices are just unbound variables (thus coindexing is unification and contraindexing is an inequality constraint), an interesting technical problem arises if the basic parsing engine uses a chart (Kay 1967, 1980). Because it is fundamental to such devices that the label on an edge is copied before it is used as a component of a larger phrase, the variables representing indices will be copied or renamed and the indices on the WH item and its sponsored trace will no longer be identical. However, it is important that the sharing of variables among the components of an ELI be respected when they come together in a phrase. One way of overcoming this problem is to associate a vector of variables with each edge, in which each variable that is shared between two or more edges is assigned a unique position. Whenever edges are combined their associated vectors are unified, thus ensuring that the corresponding variables in each edge are identified.</Paragraph>
    <Paragraph position="4"> Finally, our linguistic examples suggest to us that a more focused notion of sponsoring might be formulated. We observe that, modulo adjunction, empty nodes tend to stand in fixed structural relations to their sponsors. If this is indeed generally true, then these strong locality constraints should clearly be exploited in the parsing process. This amounts to adopting the framework of Tree Adjoining Grammars (Frank 1990; Joshi, Levy, and Takahashi 1975; Kroch and Joshi 1985; Schabes 1990). The emphasis would then fall on deriving the initial and auxiliary trees from the general principles of grammar.</Paragraph>
  </Section>
class="xml-element"></Paper>