<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-2005">
  <Title>Squibs and Discussions Parsing and Empty Nodes</Title>
  <Section position="5" start_page="292" end_page="296" type="metho">
    <SectionTitle>
4. Linguistic Aspects of Sponsoring
</SectionTitle>
    <Paragraph position="0"> The goal of this section is to demonstrate that the constraints that sponsoring imposes are plausible with respect to current linguistic assumptions. To the extent that they are, an important step will have been taken in establishing the decidability of these theories.</Paragraph>
    <Paragraph position="1"> Consider once again the example in Figure 1. Because there is a government relationship between the V trace and the NP trace, and a movement relationship between the gave node under I and the V trace, it seems reasonable to include both of these traces in the ELI that permits gave to appear under I. The alternative clearly exists of allowing every N to sponsor an NP trace to allow, for example, for heavy NP shift of its maximal projection. It does not matter that this would lead to the provision of empty NP nodes in cases where no movement could occur, because the structures produced by the parser must independently satisfy all the requirements of the grammar.</Paragraph>
    <Paragraph position="2"> Now consider an example involving cyclic WH-movement, as depicted in Figure 3.</Paragraph>
    <Paragraph position="3"> For English, WH-items such as which could sponsor the NP trace at the base of the chain (in Figure 3 the NP trace in embedded object position). However, we have already motivated a trace for the subcategorized complement, which should also serve as the foot of the WH-chain. Sponsors must also be found for the intermediate traces.</Paragraph>
    <Paragraph position="4">  Because the number of intermediate traces that can appear in WH-constructions is not bounded, these intermediate traces cannot be sponsored by either the WH-item or the embedded verb. However, they can be sponsored by the bridge verbs that govern their CP parents. For example, the intermediate NP trace in Figure 3 is sponsored by the verb think.</Paragraph>
    <Paragraph position="5"> Another possibility, inspired by the work of Grimshaw (1990) and Speas and Fukui (1986) on extended projections, is that inflection sponsors a complete set of empty functional nodes (and their specifiers) that can appear in the clause. In this example, the intermediate trace would be sponsored by the inflection -s on knows. While the first approach is perhaps more elegant, the second one also covers relative clauses, as discussed below. Either way, each potential location of an intermediate trace will have a sponsor; it is the independent constraints on WH-movement that are responsible for ensuring that, if a trace appears, it will be properly incorporated into a WH-chain.</Paragraph>
    <Paragraph position="6"> The verb movement in Figure 3 can be treated as before. Presumably the ELI for does that permits it to appear under C also sponsors the corresponding trace in I, and the ELI for knows (or perhaps for the inflectional ending -s) that permits the verb to appear under I also sponsors the trace in V.</Paragraph>
    <Paragraph position="7"> Next, consider the example in Figure 4. As well as the by now familiar V to ! movement, it also exhibits two examples of empty categories that are not members of chains, so their sponsors cannot be determined on this basis. Responsibility for sponsorship of the empty C as well as the PRO could be ascribed to the verb wants that governs the CP in which they appear. This is a control verb and is in any case responsible for identifying George as the antecedent of the PRO. According to this view the inflected verb wants (i.e., the lexical stem and the inflection) sponsors a total of three empty categories. Alternatively, one could allow the infinitive marker to to sponsor PRO and the empty complementizer. This approach is consistent with the  Empty operators in relative clauses.</Paragraph>
    <Paragraph position="8"> view that inflection sponsors all of the empty functional nodes of the clause in which it appears.</Paragraph>
    <Paragraph position="9"> English relative clauses are a major challenge for the sponsoring approach. Even though relative clauses share many of the structural properties of WH-question constructions such as cyclic movement, they can appear in a greater diversity of forms. All we attempt here is a survey of the problems encountered in developing the sponsoring account of empty nodes in relative clause constructions and their possible solutions. Consider the case of a relative clause introduced by an empty operator Op (rather than an overt relative pronoun), such as the example in Figure 5.</Paragraph>
    <Paragraph position="10"> The analyses discussed above provide sponsors for every empty node except the empty relative operator Op in the specifier position of CP. Because the number of relative clauses introduced by empty operators is not bounded (examples such as Example I seem to be indefinitely iterable) we are effectively forced to the conclusion that inflection, or some other lexical element present inside each relative clause, sponsors the empty operator Op in examples such as Example 1 and Figure 5.</Paragraph>
    <Paragraph position="11"> Example 1 A man lOp1 Kim likes hi, lOp2 Sandy hates t2\] ... and \[Op3 Ivan ignores t3\] ...</Paragraph>
    <Paragraph position="12"> Even though the structure of reduced relative clauses such as Example 1 is not as well understood as ordinary relatives, they presumably involve empty operators as well. Assuming that we analyze the passive morphology on the participle as inflection (this seems linguistically motivated if we assume that movement to subject position is A-movement), the empty operator and all of its traces would be appropriately licensed.</Paragraph>
    <Paragraph position="13">  A horse \[cpOp\[ipt\[vp ridden t\] past the barn \]\] fell.</Paragraph>
    <Paragraph position="14"> Finally, relative clauses can extrapose quite freely, as in Figure 6. (This diagram assumes that extraposed relative clauses adjoin to IP, but nothing rests on this assumption.) null The sponsoring mechanisms discussed above account for all of the empty nodes except for the trace of the extraposed relative clause (adjoined to N ~ in Figure 6). As an anonymous Computational Linguistics reviewer points out, an apparently unbounded number of relative clauses can be extraposed from a single NP.</Paragraph>
    <Paragraph position="15"> Example 3 A \[N'\[N'\[N' photo tl\]t2\]t3\] appeared in today's paper \[CP3 taken by Mapplethorpe\] \[cP2 showing him smiling\]... \[cp, that I think you would like\].</Paragraph>
    <Paragraph position="16"> Just as in the case of empty operator relative clauses, this iterability suggests that the extraposition traces must be sponsored by lexical items that appear inside the extraposed relative clauses. The inflection element in the relative clause seems to be the most appropriate lexical item to sponsor these traces.</Paragraph>
    <Paragraph position="17"> To summarize, it seems that it is possible to identify sponsors for the empty nodes for a variety of linguistic constructions. Of course, the above examples do not demonstrate that it will always be possible to identify appropriate sponsors. In any given theory a detailed analysis of each construction is required to determine the appropriate sponsors.</Paragraph>
  </Section>
  <Section position="6" start_page="296" end_page="297" type="metho">
    <SectionTitle>
5. Implementation
</SectionTitle>
    <Paragraph position="0"> The diagram in Figure 7 shows the structure of a possible implementation of a parsing system that exploits the notion of sponsoring. Square cornered boxes are used for data, and round corners for processes. Lexical access is applied to the input string to produce (nondeterministically) the extended lexical item (ELI) of each word. Its output is split into a sequence of lexical items and a bag of empty nodes.</Paragraph>
    <Paragraph position="1"> The parser can be based on any standard algorithm. It is special only in that all the terminal items in the phrases it constructs come either from the string of lexical items or from the bag of empty nodes, so it is impossible for any empty node to appear more than once in an analysis.</Paragraph>
    <Paragraph position="2"> An obvious defect of the architecture in this simple form is that, in the absence of some form of prediction, the parser will consider at all points in the string all the structures that can be built entirely from empty nodes. A simple solution to this problem is to compute all the trees consisting solely of empty nodes sponsored by lexical items appearing in the utterance to be parsed before beginning the main parsing process. This will make it possible to use a parser that does not deal with empty nodes explicitly. The idea is to modify the interface between the parser and the grammar.</Paragraph>
    <Paragraph position="3"> The fact that sponsoring can be implemented entirely within the &amp;quot;rule-maker&amp;quot; interface shows that it can be used with any parsing algorithm. We take it that the main job of this interface will be to manufacture &amp;quot;rules&amp;quot; that enshrine the local well-formedness constraints on individual nodes. The modification consists in adding rules to this set.</Paragraph>
    <Paragraph position="4"> A rule a ~ bl ... bn will be passed in its original form to the parser, which can use it to build a phrase from n nonempty daughters. In addition, the rule maker supplies the parser with rules derived from this one by replacing k &lt; n of the bi with empty trees, yielding a rule with n-k items on the right-hand side. The parser treats all these rules on equal footing. Apart from the sponsoring relationship, there is no requirement that any of the k empty trees be related to the n-k nonempty trees that the parser proper gets to see.</Paragraph>
    <Paragraph position="5"> There are certain bookkeeping tasks that can best be undertaken by the rule maker.</Paragraph>
    <Paragraph position="6"> The most important of these have to do with ensuring that no empty terminal appears more than once in any structure. Concomitantly, it makes it possible to verify, at the end of the parse, that all empty terminals appear in the structure, if this is a requirement.</Paragraph>
    <Paragraph position="7"> The rule maker can also be used to percolate constraints up or down the tree, possibly discharging some of them in the process.</Paragraph>
    <Paragraph position="8"> One remaining problem is to arrange for the parser to include the empty trees in the structures that it builds. We assume that this information accompanies the  A simple architecture.</Paragraph>
    <Paragraph position="9">  The modified architecture.</Paragraph>
    <Paragraph position="10"> rule, perhaps in the form of features on the parent category, as in many unification grammars.</Paragraph>
  </Section>
class="xml-element"></Paper>