<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-2005">
  <Title>Squibs and Discussions Parsing and Empty Nodes</Title>
  <Section position="4" start_page="290" end_page="292" type="intro">
    <SectionTitle>
3. Sponsoring
</SectionTitle>
    <Paragraph position="0"> Our solution to this problem is a device inspired by the notion of licensing in GB (Abney 1986). According to this conception, the presence and location of each empty node is justified by the specific structural relations it stands in with other nodes. For example, every noun phrase might be required to receive Case and a &amp;role, and it may be that the phrase would have to appear at different places in the structure for both of these assigrunents to be made. However, it may also be that the phrase can be represented in one, or both, positions by a related empty category, a trace of the phrase, which is licensed by its fulfillment of this specific role.</Paragraph>
    <Paragraph position="1"> To guarantee that only a finite number of empty nodes is posited in any analysis, we propose that, whatever parsing strategy is used, there be a global constraint on the number of empty nodes that can be posited in any single search path. We require that every empty node be sponsored by some lexical or morphological item that appears in the input. By sponsoring we mean that every empty node is associated with some nonempty lexical item, which we call its sponsor, and that the number of empty nodes that a single lexical token can sponsor is fixed by the lexicon, so that the set of all empty nodes to appear in the parse can be determined directly by a simple inspection of the lexical items in the input string.</Paragraph>
    <Paragraph position="2"> Sponsorship is closely related to lexicalization in TAGs and CFGs (Schabes 1990, 1992; Schabes, AbeillG and Joshi 1988; Schabes and Waters 1993; Vijay-Shanker and Schabes 1992). In a lexicalized grammar every node in the parse tree originates from some lexical entry, so parsing becomes a jigsaw puzzle-like problem of finding a consistent way of assembling the pieces of trees associated with each lexical item.</Paragraph>
    <Paragraph position="3"> Sponsoring is a weaker notion, in that only some of the constituent structure, namely the lexical items and empty nodes, are specified in lexical entries. This seems plausible in a framework in which general principles of grammar (e.g., X ~ theory, Case theory, etc.) determine the overall structure of the parse tree. In addition, finding an appropriate association of constituent structure nodes with lexical items can be a difficult task. Because the sponsoring approach is only concerned with empty nodes, it should be easier to apply it to a wider variety of grannnatical theories than a lexicalization  The structure of the lexicon and ELIs.</Paragraph>
    <Paragraph position="4"> approach, which requires that every node (empty or not) be associated with a lexical item (but see the remarks in the conclusion below).</Paragraph>
    <Paragraph position="5"> We now discuss one way to formulate a sponsoring approach. A lexical item and the set of empty categories that it sponsors constitute an extended lexical item (ELI) as sketched in Figure 2. In simple systems, such as the parser described in the next section, each lexical and morphological entry explicitly specifies the traces that it sponsors, but in more sophisticated implementations this could be determined automatically from principles of the grammar and properties of the lexical entry. It is not intended that sponsoring be used to change grammar, but only to impose relatively weak global constraints on the appearance of empty categories.</Paragraph>
    <Paragraph position="6"> There are several variants of the basic idea. For example, one could require that all the empty nodes supplied by the lexicon be used in the analysis. On the one hand, this could lead to a proliferation of lexical entries. On the other, it could prune the search space more effectively if the role of each empty node were made as specific as possible.</Paragraph>
    <Paragraph position="7"> As we remarked, previous proposals have made the number of empty nodesposited a function of the length of the input string. The novelty of our proposal is twofold. First, it provides a finer way of estimating the number of empty nodes that will occur. In fact, in the simplest version of the theory, the number of empty and nonempty terminals in a sentence is simply the sum of the sizes of the ELIs of the words in it. The number of empty categories is therefore this number minus the number of words. The fact that the number of empty nodes is bounded before parsing begins is the most important part of our proposal.</Paragraph>
    <Paragraph position="8"> Our second proposal is that each of the items in an ELI is marked to show the specific role it must fill. Only one member, for example, will be capable of receiving a 0-role, and this member will not be capable of filling any position in which a 0-role is not assigned.</Paragraph>
  </Section>
class="xml-element"></Paper>