<?xml version="1.0" standalone="yes"?>
<Paper uid="P92-1002">
  <Title>AN ALGORITHM FOR VP ELLIPSIS</Title>
  <Section position="4" start_page="9" end_page="9" type="metho">
    <SectionTitle>
THE ALGORITHM
</SectionTitle>
    <Paragraph position="0"> The input to the algorithm is an elliptical VP(VPE), and VPlist, a list of VP's occurring in the current sentence, and those occurring in the two immediately preceding sentences. In addition, it is assumed that the parse trees of these sentences are available as global variables, and that NP's in these parse trees have been assigned indices to indicate coreference and quantifier binding.</Paragraph>
    <Paragraph position="1"> The antecedent selection function is:</Paragraph>
    <Paragraph position="3"> First, impossible antecedents are removed from the VPlist. Then, the remaining items in VPlist are assigned preference levels, and the item with the highest preference level is selected as the antecedent. If there is more than one item with the same preference level, the item closest to the VPE, scanning left from the VPE, is selected.</Paragraph>
    <Paragraph position="4"> The definition of the function remove-impossible is as follows: remove-impossible(VPlist,VPE) For all v in VPlist if ACD(v,VPE) or BE-DO-conflict(v,VPE) then remove(v, VPlist) There are two types of impossible antecedents: the first involves certain antecedent-containment structures, and the second involves cases in which the antecedent contains a BE-form and the target contains a DO-form. These are described in detail below.</Paragraph>
    <Paragraph position="5"> Next, preference levels are assigned to remaining items in VPlist by the assign-levels function. (All items on VPlist are initialized with a level of 0.) assign-levels (VPlist, VPE) For all v in VPlist if related-clause(v,VPE) then</Paragraph>
    <Paragraph position="7"> An antecedent is preferred if there is a clausal relationship between its clause and the VPE clause, or if the antecedent and the VPE have coreferential subjects. The determination of these preferences is described in detail below.</Paragraph>
    <Paragraph position="8"> Finally, the select-highest function merely selects the item on VPlist with the highest preference level. If there is more than one item with the highest preference level, the item nearest to the VPE (scanning left) is selected.</Paragraph>
  </Section>
  <Section position="5" start_page="9" end_page="9" type="metho">
    <SectionTitle>
IMPOSSIBLE ANTECEDENTS
</SectionTitle>
    <Paragraph position="0"> This section concerns the removal of impossible antecedents from VPlist. There are two cases in which a given VP is not a possible antecedent.</Paragraph>
    <Paragraph position="1"> The first deals with antecedent-containment, the second, with conflicts between BE-forms and DOforms. null</Paragraph>
  </Section>
  <Section position="6" start_page="9" end_page="10" type="metho">
    <SectionTitle>
ANTECEDENT CONTAINMENT
</SectionTitle>
    <Paragraph position="0"> There are cases of VP ellipsis in which the VPE is contained within the antecedent VP: IV \[... VPE ...\]\]vP Such cases are traditionally termed antecedent-contained deletion (ACD). They are highly constrained, although the proper  formulation of the relevant constraint remains controversial. It was claimed by May (1985) and others that ACD is only possible if a quantifier is present. May argues that this explains the following contrast: (2) a. Dulles suspected everyone who Angelton did.</Paragraph>
    <Paragraph position="1"> b. * Dulles suspected Philby, who Angelton did.</Paragraph>
    <Paragraph position="2"> However, it has been subsequently noted (cf. Fiengo and May 1991) that such structures do not require the presence of a quantifier, as shown by the following examples:  (3) a. Dulles suspected Philby, who Angelton did too.</Paragraph>
    <Paragraph position="3"> b. Dulles suspected Philby, who Angelton  didn't.</Paragraph>
    <Paragraph position="4"> Thus the algorithm will allow cases of ACD in which the target is dominated by an NP which is an argument of the antecedent verb. It will not allow cases in which the target is dominated by a sentential complement of the antecedent verb, such as the following: (4) That still leaves you a lot of latitude. And I suppose it did.</Paragraph>
    <Paragraph position="5"> Here, &amp;quot;suppose&amp;quot; is not a possible antecedent for the elliptical VP. In general, configurations of the following form are ruled out:</Paragraph>
  </Section>
  <Section position="7" start_page="10" end_page="10" type="metho">
    <SectionTitle>
IV \[... VPE .--\]s-..\]vP
BE/DO CONFLICTS
</SectionTitle>
    <Paragraph position="0"> The auxiliary verb contributes various teatures to the complete verb phrase, including tense, aspect, and polarity. There is no requirement that these features match in antecedent and elliptical VP.</Paragraph>
    <Paragraph position="1"> However, certain conflicts do not appear to be possible. In general, it is not possible to have a DO-form as the elliptical VP, with an overt BE-form in the antecedent. Consider the following example: (5) Nor can anyone be certain that Prokofief would have done better, or even as well, under different circumstances. His fellowcountryman, Igor Stravinsky, certainly did not.</Paragraph>
    <Paragraph position="2"> In this example, there are two elements on the VP list: &amp;quot;be certain...&amp;quot;, and &amp;quot;do better&amp;quot;. The target &amp;quot;did not&amp;quot; rules out &amp;quot;be certain&amp;quot; as a possible antecedent, allowing only the reading &amp;quot;Stravinsky did not do better&amp;quot;. If the elliptical VP is changed from &amp;quot;did not&amp;quot; to &amp;quot;was not&amp;quot;, the situation is reversed; the only possible reading is then &amp;quot;Stravinsky was not certain that Prokofief would have done better...&amp;quot;.</Paragraph>
    <Paragraph position="3"> A related conflict to be ruled out is that of ac- null tive/passive conflicts. A passive antecedent is not possible if the VPE is a DO-form. For example: (6) Jubal did not hear of Digby's disappear null ance when it was announced, and, when he did, while he had a fleeting suspicion, he dismissed it; In this example, &amp;quot;was announced&amp;quot; is not a possible antecedent for the VPE &amp;quot;did&amp;quot;.</Paragraph>
    <Paragraph position="4"> One possible exception to this rule involves progressive antecedents, which, although they contain a BE-form, may be consistent with a DO-form target. The following (constructed) example seems marginally acceptable: (7) Tom was cleaning his room today. Harry did yesterday.</Paragraph>
    <Paragraph position="5"> Thus a BE-form together with a progressive does not conflict with a DO-form.</Paragraph>
  </Section>
  <Section position="8" start_page="10" end_page="10" type="metho">
    <SectionTitle>
PREFERENCE LEVELS
</SectionTitle>
    <Paragraph position="0"> If there are several possible antecedents for a given VPE, preferences among those antecedents are determined by looking for other relations between the VPE clause and the clauses containing the possible antecedents.</Paragraph>
  </Section>
  <Section position="9" start_page="10" end_page="10" type="metho">
    <SectionTitle>
CLAUSAL RELATIONSHIPS
</SectionTitle>
    <Paragraph position="0"> An antecedent for a given VPE is preferred if there is a configurational relationship between the antecedent clause and the VPE clause. These include comparative structures and adverbial clauses.</Paragraph>
    <Paragraph position="1"> Elliptical VP's (VPE) in comparative constructions are of the form</Paragraph>
  </Section>
  <Section position="10" start_page="10" end_page="11" type="metho">
    <SectionTitle>
\[VP Comparative \[NP VPE\]\]
</SectionTitle>
    <Paragraph position="0"> where Comparatives are expressions such as &amp;quot;as well as&amp;quot;, &amp;quot;better than&amp;quot;, etc. In constructions of this form there is a strong preference that VP is the antecedent for VPE. For example: (8) Now, if Morton's newest product, a corn chip known as Chip-o's, turns out to sell as well as its stock did...</Paragraph>
    <Paragraph position="1"> Here, the antecedent of the VPE &amp;quot;did&amp;quot; is the VP &amp;quot;sell&amp;quot;.</Paragraph>
    <Paragraph position="2"> The next configuration involves VPE's within adverbial clauses. For example, (9) But if you keep a calendar of events, as we do, you noticed a conflict.</Paragraph>
    <Paragraph position="3"> Here the antecedent for the VPE &amp;quot;do&amp;quot; is &amp;quot;keep a calendar of events&amp;quot;. In general, in configurations of the form:</Paragraph>
  </Section>
  <Section position="11" start_page="11" end_page="11" type="metho">
    <SectionTitle>
\[VP ADV \[NP VPE\]\]
</SectionTitle>
    <Paragraph position="0"> VP is preferred over other possible antecedents. null It is important to note that this is a preference rule, rather than an obligatory constraint. Although no examples of this kind were found in the Brown Corpus, violations of this constraint may well be possible. For example: (10) John can walk faster than Harry can run. Bill can walk faster than Barry can.</Paragraph>
    <Paragraph position="1"> If a reading is possible in which the VPE is &amp;quot;Barry can run&amp;quot;, this violates the clausal relationship preference rule.</Paragraph>
  </Section>
  <Section position="12" start_page="11" end_page="11" type="metho">
    <SectionTitle>
SUBJECT COREFERENCE
</SectionTitle>
    <Paragraph position="0"> Another way in which two clauses are related is subject coreference. An antecedent is preferred if its subject corefers with that of the elliptical VP.</Paragraph>
    <Paragraph position="1"> An example: (11) He wondered if the audience would let him  finish. They did.</Paragraph>
    <Paragraph position="2"> The preferred reading has &amp;quot;they&amp;quot; coreferential with &amp;quot;the audience&amp;quot; and the antecedent for &amp;quot;did&amp;quot; the VP &amp;quot;let him finish&amp;quot;.</Paragraph>
    <Paragraph position="3"> Subject &amp;quot;coreference&amp;quot; is determined manually, and it is meant to reflect quantifier binding as well as ordinary coreference - that is, standard instances involving coindexing of NP's.</Paragraph>
    <Paragraph position="4"> Again, it must be emphasized that the subject coreference rule is a preference rule rather than an obligatory constraint. While no violations were found in the Brown corpus, it is possible to construct such examples.</Paragraph>
  </Section>
  <Section position="13" start_page="11" end_page="11" type="metho">
    <SectionTitle>
INTERACTION OF PREFERENCE
RULES
</SectionTitle>
    <Paragraph position="0"> There are cases where more than one preference rule applies. The antecedent selected is the item with the highest preference level. If more than one item has the same preference level, the item nearest to the VPE is selected, where nearness is determined by number of words encountered scanning left from the VPE.</Paragraph>
    <Paragraph position="1"> In the following example, two preference rules apply: (12) usually, this is most exasperating to men, who expect every woman to verify their preconceived notions concerning her sex, and when she does not, immediately condemn her as eccentric and unwomanly.</Paragraph>
    <Paragraph position="2"> The VPE clause is an adverbial clause modifying the following clause. Thus the VP &amp;quot;condemn her as eccentric and unwomanly&amp;quot; receives a preference level of 1. The subject &amp;quot;she&amp;quot; of the VPE is coindexed with &amp;quot;every woman&amp;quot;. This causes the VP &amp;quot;verify their preconceived notions concerning her sex&amp;quot; to also receive a preference level of 1. Since both of these elements have the same preference level, proximity is determined by scanning left from the VPE. This selects &amp;quot;verify their preconceived notions concerning her sex&amp;quot; as the antecedent. null</Paragraph>
  </Section>
  <Section position="14" start_page="11" end_page="11" type="metho">
    <SectionTitle>
TESTING THE ALGORITHM
</SectionTitle>
    <Paragraph position="0"> The algorithm has been tested on a set of 304 examples of VP ellipsis collected from the Brown Corpus. These examples were collected using the UNIX grep pattern-matching utility. The version of the Brown Corpus used has each word tagged by part of speech. I defined search patterns for auxiliary verbs that did not have verbs nearby. These patterns did not succeed in locating all the instances of VP ellipsis in the Brown Corpus. However, the 304 examples do cover the full range of types of material in the Brown Corpus, including both &amp;quot;Informative&amp;quot; (e.g., journalistic, scientific, and government texts) and &amp;quot;Imaginative&amp;quot; (e.g., novels, short stories, and humor). I have divided these examples into three categories, based on whether the antecedent is in the same sentence as the VPE, the adjacent (preceding) sentence, or earlier (&amp;quot;Long-Distance&amp;quot;). The definition of sentence is taken from the sentence divisions present in the Brown Corpus.</Paragraph>
  </Section>
class="xml-element"></Paper>