<?xml version="1.0" standalone="yes"?>
<Paper uid="E95-1014">
  <Title>Corpus-based Method for Automatic Identification of Support Verbs for Nominalizations</Title>
  <Section position="5" start_page="99" end_page="101" type="metho">
    <SectionTitle>
4 Experiment with
</SectionTitle>
    <Paragraph position="0"> appeal-appeal We have taken for example the case of the verb appeal which was interesting since its corresponding deverbal noun shares the same surface form appeal. In order to extract corpus evidence, we used a lexical transducer of English that, given the surface word appeal, produced all the inflected forms appeal, appeal's, appealing, appealed, appeals and appeals '.</Paragraph>
    <Paragraph position="1"> Using these surface forms as a filter, we scanned 134 Megabytes of tokenized Associated Press newswire stories from the year 19895. As a result of filtering, 6704 sentences (1 Mbyte of text) were extracted. This text was part-of-speech tagged using the Xerox HMM tagger (Cutting et al., 1992). The lexical entries corresponding to appeal were tagged with the following tags: as a noun (3910 times), as an active or infinitival verb (1417), as a progressive verb (292), and as a past participle (400).</Paragraph>
    <Paragraph position="2"> This tagged text was then parsed by a low-level dependency parser (Grefenstette, 1994)\[Chap 3\]. From the output of the dependency parser we extracted all the lexically normalized verbs of which appeal was tagged as a direct object. The most common of these verbs are shown in Figure 2.</Paragraph>
    <Paragraph position="3"> Our speaker's intuition tells us that the support verb for the nominalized use of appeal is make. But this data does not give us enough information to make this judgement, since concrete versions as a separate entity are not distinguishable from nomlnalizations of the verb.</Paragraph>
    <Paragraph position="4"> In order to separate nominalized uses of the predicate appeal from concrete uses, we will refer to the linguistic discussion presented in the introduction that says that nominalizations retain some of the argument/adjunct structure of the verbal predicate. This is verified in the corpus since we find many parallel SThis corresponds to 20 million words of text. structures involving appeal both as a verb and as a noun, such as:</Paragraph>
    <Section position="1" start_page="99" end_page="101" type="sub_section">
      <SectionTitle>
Vice President Salvador Laurel said
</SectionTitle>
      <Paragraph position="0"> today that an ailing Ferdinand Marcos may not survive the year and appealed to President Corazon Aquino to allow her ousted predecessor to die in his homeland.</Paragraph>
      <Paragraph position="1"> Mrs. Marcos made a public appeal to President Corazon Aqulno to allow Marcos to return to his homeland to die.</Paragraph>
      <Paragraph position="2"> Indeed, if we examine a common nominalization transformation, i.e. that of transforming the direct object of a verb into a Norman genitive of the nominalized form, we find a great overlap in the lexical arguments 6.  The parser's output allowed us to extract patterns involving prepositional phrases following noun phrases headed by appeal as well as those following verb sequences headed by appeal. The most common prepositional phrases found after appeal as a verb began with the prepositions7: Lo (466 times), for (145), in (18), on (12), wilh (5), etc. The prepositional phrases following appeal as a noun are headed by to (321 times), for (253), in (200), of (134), from (78), on (34), etc. The correspondence between the most frequent prepositions allowed us to consider that the patterns of a noun phrase headed by appeal followed by one these prepositional phrases (i.e., begun with to, for, and in) constituted true nominalizations s. There were 6We decided not to use this type of data in our experiments because matching lexical arguments requires much larger corpora than the ones we had extracted for the other verbs tested.</Paragraph>
      <Paragraph position="3"> rWe ignored prepositionM phrases headed by by as being probable passivizations, since our parser does not recognize passive patterns involving by.</Paragraph>
      <Paragraph position="4"> SHere we used only part of the corpus evidence that was available. Other patterns of nominMizations of appeal, e.g. Saxon genitives like the criminal's appeal, may well exist in the corpus.  structure NP PP where 'appeal' heads the NP and where one of {to, for, in} begins the PP. 774 instances of these patterns.</Paragraph>
      <Paragraph position="5"> The parser's output further allowed us to extract the verbs for which these nominalizations were considered as the direct objects. 318 of these nominal syntactic patterns including to, for and in were found. Of these patterns, the main verb supporting the objective nominalizations are shown in Figure 3. These results suggest that the support verb for the nominalization of appeal is make.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="101" end_page="101" type="metho">
    <SectionTitle>
5 Other Predicate Examples
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="101" end_page="101" type="sub_section">
      <SectionTitle>
and Discussion
</SectionTitle>
      <Paragraph position="0"> When the same filtering technique is applied to subcorpora derived for other nominalization pairs, we obtain the results given in Figure 4. For each verb-noun pair all sentences containing any form of the words were extracted from the AP corpus. The sentences were processed as explained in section 3. For each verbal use, the most frequent prepositional phrases following the verbs were tabulated and the three most frequent prepositions were retained. For example, the most frequent prepositions beginning prepositional phrases following verb uses of the lemma offer were for, in and to. These prepositions were used to select probable nominalizations by extracting noun uses of the predicates that were immediately followed by prepositional phrases headed by one of the three most frequent verbal prepositions. For these extracted noun phrases, when they were found in a direct object position, the main verbs were tabulated which gives the results in Figure 4.</Paragraph>
      <Paragraph position="1"> Some of the results in Figure 4 correspond to our naive intuitions of collocational support verbs, such as make an offer. For discussion, both have and hold appear equally frequently. But other words show the limitations of this method, we would expect make a demand where we find meet a demand. In the same subcorpus, although we find make a demand 77 times, meet a demand is twoand-a-half times more common. Could this be because, in a newswire corpus, meeting a demand is more newsworthy than making one? If we just look at the cases where demand is modified by the indefinite article; which might correspond to the more generic nominalizations one spontaneously creates when generating examples, we find that in the corpus make a demand occurs slightly more often than meet a demand, ten times vs. six times, but this is too rare to use as a criterion.</Paragraph>
      <Paragraph position="2"> In other cases, such as with proposal and assertion we find make and reject with almost equal frequency, and though make might well be considered a support verb, it is hard to accept reject as semantically empty. Though reject is more a consequence than an antonym of make a proposal, this raises the question, to which we have no answer, of whether support verbs have an equally empty antonym.</Paragraph>
      <Paragraph position="3"> A more interesting case is the appearance of issue for order and warning where we would expect give. Looking into the corpus evidence, we find issue a restraining order 46 times, and give any type of order only 16 times. This evidence suggest a limitation of our word-based approach. Multi-word phrases, such as the nominalized phrase restraining order, might take a different support verb than the simple unqualified word forms, such as order.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>