<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0702">
  <Title>I I I I I I I I I I I I I I I I I I I Disambiguating Verbs with the WordNet Category of the Direct Object</Title>
  <Section position="4" start_page="0" end_page="10" type="metho">
    <SectionTitle>
3 Aspectually Ambiguous Verbs
</SectionTitle>
    <Paragraph position="0"> While some verbs appear to connote only one aspectual class regardless of context, e.g., stare (nonc-lminated event), many verbs are aspectually am.</Paragraph>
    <Paragraph position="1"> biguous. For example, shaw denotes a state in, H/$ lumbar puncture showed evidence of white cells, but denotes an event in, He showed me the photographs.</Paragraph>
    <Paragraph position="2"> This ambiguity presents a di~culty for automatically classifying a verb because the aspectual class of a clause is a function of several clausal constituents in addition to the main verb (Dowry, 1979; Moens and Steedman, 1988; Pustejovsky, 1991). However, previous work that numerically evaluates aspectual classification has looked at verbs in isolation (Klavans and Chodorow, 1992; Siegel, 1997).</Paragraph>
    <Paragraph position="3">  The verb have is particularly problematic. In the medical domain, have occurs as the main verb of clauses frequently (8% of clauses) and is aspectually ambiguous, occurring 69.9% of the time as a state, and 30.1% of the time as an event. Most other ambiguous verbs are more highly dominated by one sense in this domain (Siegel, 1998).</Paragraph>
    <Paragraph position="4"> In this section, I examine factors contributing to aspectual ambiguity. First, I exam the interaction between a verb and its arguments in determining aspectual class. The semantic category of open class words plays a large role in this process. And second, I describe a semantic hierarchy of statively ambiguous verb. This hierarchy groups together verbs that tend to interact with their arguments in similar ways.</Paragraph>
    <Section position="1" start_page="10" end_page="10" type="sub_section">
      <SectionTitle>
3.1 How Clausal Constituents Contribute
to Aspectual Class
</SectionTitle>
      <Paragraph position="0"> The presence, syntactic categories, lexical heads, and plurality of a verb's arguments influence aspectual class. This is illustrated in Table 2, which shows example clausal features that influence aspectual class. The effect of each feature is illustrated by showing two similar sentences with distinct aspectual classes.</Paragraph>
      <Paragraph position="1"> The number of ways in which clausal constituents interactively influence aspect is ,mknown. However, syntax alone is not sufficient, and the lexical head of multiple constituents (e.g., the verb phrase and the direct object) are often factors. Moreover, the semantic category of these features can also play a role. For example, Sue played the piano is nonc,lminated, while Sue played the sonata signifies a c-lminated event (this example comes from Moens and Steedman (1988)).</Paragraph>
    </Section>
    <Section position="2" start_page="10" end_page="10" type="sub_section">
      <SectionTitle>
3.2 Classes of Ambiguous Verbs
</SectionTitle>
      <Paragraph position="0"> Placing aspectually ambiguous verbs into semantic categories will help predict how these verbs combine with their arguments to determine aspectual class. This is because many verbs with related meanings combine with their arguments in similar ways.</Paragraph>
      <Paragraph position="1"> In general, there is a correlation between a verb's subcategorization frame and semantic class (Levin, 1993), and this applies to aspect in particular.</Paragraph>
      <Paragraph position="2"> For example, look and weigh can each appear as events, e.g., I looked at the baby. (event) I we/ghed the baby. (event) and can also appear as states, as in, The baby ~ heavy. (state) The baby weighed a lot. (state) Is this illustrates, these two verbs have similar sub-categorization frames that determine their aspectual class. There is also a relationship between their meanings, since each describes a type of perception or measurement.</Paragraph>
      <Paragraph position="4"> John drove the car. P John drove the car. P John saw Sue. P Judith p/ayed the piano. P John ate fries. P Kathy sho~ed people her car. P Kathy showed people her car. P Judith looked around the store. P Kathy shot at deer. P Sal said that it helps. C Contrasting Example: John drove the car ragged. John drove the car up.</Paragraph>
      <Paragraph position="5"> John saw that Sue was happy. Judith p/ayed the sonata.</Paragraph>
      <Paragraph position="6"> John ate the fries.</Paragraph>
      <Paragraph position="7"> Kathy shorted the people her car. Kathy showed Sal her car.</Paragraph>
      <Paragraph position="8"> Judith looked around the corner. Kathy shot at the deer.</Paragraph>
      <Paragraph position="9"> Sal says that it helps.</Paragraph>
      <Paragraph position="10">  Example verbs: admit, confirm, indicate, say judge, remember, think, wish feel, see, smell, weigh astonish, dismay, please, surlmdse hold, lie, sit, stand work, run continue , remain Event sentence: I said, #Hello.&amp;quot; I thought about them.</Paragraph>
      <Paragraph position="11"> I felt the tablee/oth.</Paragraph>
      <Paragraph position="12"> You surprised me.</Paragraph>
      <Paragraph position="13"> I lay on the bed.</Paragraph>
      <Paragraph position="14"> I worked hard.</Paragraph>
      <Paragraph position="15"> \[ continued to talk about it. \[ State sentence: \[ say it is correct.</Paragraph>
      <Paragraph position="16"> I think they are nzce.</Paragraph>
      <Paragraph position="17"> I felt terrible.</Paragraph>
      <Paragraph position="18"> That suprises me.</Paragraph>
      <Paragraph position="19">  The book lies on the bed. The machine works.</Paragraph>
      <Paragraph position="20"> I continued to feel good. Table 3 shows the top level of a hierarchy of star tively ambiguous verbs. Seven semantic groups are shown, each with a set of example verbs, and two sentences illustrating contrasting uses of an example verb fxom that group. Each verb in the first group, communication, can appear as either an event or state. Intuitively, this is because each verb can convey a communicative act, e.g., She s.howed me the photos. (event) or, alternatively, a non-dynamic situation, e.g., The zrays show no sign ol ~rth. (state) Verbs in the second group in Table 3, cognitive, can convey a mental event, e.g., When he mentioned bananas, she remembered Edward. (event) or, alternatively, a mental state, e.g., I'U ahvays remember Disney WorlcL (state) The groups perception and psych-movement are subgroups of cognition. The perception and communication groups have previously been idestiffed with respect to aspect in particular (Vendler, 1967; Dowry, 1979), and those and psych-movement for general purposes beyond aspectual ambiguity (Levin, 1993). The fifth group, locative, has previously been identified as &amp;quot;lay-verbs. ~ (Dowty, 1979) The group metaphorical in Table 3 contains event verbs with idiomatic uses that are stative. These idiomatic uses correspond to a metaphorical interpretation of the event reading (Alexander D. Chaifee, personal communication). For example, I ra_.nn down the street. (event) It runs in the family. (state) Finally, cart/st verbs simply reflect the aspectual class of their clausal argument.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="10" end_page="11" type="metho">
    <SectionTitle>
4 Corpus: Medical Reports
</SectionTitle>
    <Paragraph position="0"> Our experiments are performed across a corpus of 3,224 medical discharge summaries comprised of 1,159,891 words. A medical discharge s,,mmary describes the symptoms, history, diagnosis, treatment and outcome of a patient's visit to the hospital. Aspectual classification is necessary for several medical report processing tasks, since these reports describe events and states that progress over time (Friedman et al., 1995).</Paragraph>
    <Paragraph position="1"> These reports were parsed with the EngLish Slot Grammar (McCord, 1990), resulting in 97,973 clauses that were parsed fully with no self-diagnostic errors (error messages were produced on some of this corpus' complex sentences). Parsing is needed to identify the main verb and direct object of each clause, as well as the presence of aspectual markers for related statistical work, described below in Section 7.</Paragraph>
    <Paragraph position="2"> Be and have are the two most popular verbs, covering 31.9% of the clauses in this corpus. Clauses with be as their main verb, composing 23.9% of the corpus, always denote a state. Clauses with have as their main verb, composing 8.0% of the corpus, are statively ambiguous. In this domain, most clauses with main verbs other than be and have can be aspectually classified by the the main verb only, e.g., by</Paragraph>
    <Paragraph position="4"> using numerical linguistic indicators (Siegel, 1998) In order to produce supertrised data with which to develop and evaluate our approach, a batch of 206 have-clauses f~om the parsed corpus were manually marked according to stativity. As a linguistic test for marking, each clause was tested for readability with, What happened was... In a separate study, a comparison between two human markers using this test to classify clauses over all verbs showed an agreement of approximately 91% (Siegel, 1998).</Paragraph>
    <Paragraph position="5"> The marked clauses, divided equally into training and testing sets of 103 clauses each, were used to develop and evaluate our approach, respectively.</Paragraph>
  </Section>
  <Section position="6" start_page="11" end_page="12" type="metho">
    <SectionTitle>
5 Applying WordNet
</SectionTitle>
    <Paragraph position="0"> I have manually designed a rule for classifying have-clauses according to stativity by the WordNet category of the direct object. To design this rule, the following were observed:  * Distributions of objects of have over the corpus. * Linguistic intuition regarding WordNet categories and aspectual class.</Paragraph>
    <Paragraph position="1"> * Correlations between the WordNet category of the direct object and stativity over the super null vised training data.</Paragraph>
    <Paragraph position="2"> To accumulate this information, WordNet was queried for each direct object of the parsed corpus. In particular, each noun was placed into one of the 25 categories at the top of WordNet's semantic hierarchy, listed in Table 4. Many nouns have multiple entries corresponding to multiple senses. As an initial approach, we take the first WordNet category listed, i.e., the most f~equent sense. Pronouns such as him and it were assigned their own category, pronoun.</Paragraph>
    <Paragraph position="3"> As shown, in Table 5, the most frequent objects of have are primarily specific to the medical domain. This table shows the high level semantic category assigned by WordNet and the classification of have-clauses with each noun as a direct object. WordNet is able to handle this technical domain since 89.1% of have-clauses have direct objects that are widelyknown medical terms and non-technical terms. The rule shown in Table 6 classifies have-clauses based on the semantic category of their direct object. In particular, clauses with direct objects that belong to the categories event, act, phenomenon, communication, possession and food are classified as events, and all others are classified as states. Linguistic insights guided the design of this rule. For example, if the direct object of have denotes an event, such as seizure, the clause describes an event. For this reason, it is clear why the WordNet categories event, act, phenomenon and communication each indicate an event clause. Note that nominalized event verbs, e.g., resolution, are placed in these four categories by WordNet. The category possession  category, and the aspectual class of have-clauses with the object. Asterisks (*) denote classifications that were intuitively derived, since these examples did not occur in the training cases.</Paragraph>
    <Paragraph position="4"> was selected since, as shown in Table 6, most occurrences of possession as a direct object of have are instances of loss, e.g., The patient had blood loss describes an event. The category food was selected to cover idioms such as The patient had lunch (event). Furthermore, this classification rule is quantitatively supported over the supervised training data.</Paragraph>
    <Paragraph position="6"/>
    <Paragraph position="8"> corpus, from which the supervised training and testing data were extracted.</Paragraph>
    <Paragraph position="9"> For each WordNet category, Table 4 shows the distribution of event and stative have-clauses with a direct object belonging to that category. As shown, each WordNet category llnimd to states with our rule occurs at least as frequently in stative clauses as they do in event clauses within the training set, with the exception of coltmication, possession and food.</Paragraph>
    <Paragraph position="10"> However, these categories occur only one time each in the training data, which is too sparse to counter linguistic intuition.</Paragraph>
  </Section>
  <Section position="7" start_page="12" end_page="13" type="metho">
    <SectionTitle>
6 Results
</SectionTitle>
    <Paragraph position="0"> There is a strong correlation between the Word-Net category of a direct object, and the aspectual class of have-clauses it appears in. When using the classification rule established in the previous subsection, the WordNet categories that appear more than five times in the supervised test data correctly predict the class of have-clauses with an average precision of 82.7deg/o. Specifically, act and event predict event have-clauses 85.7% and 66.7% correctly, respectively, and states are predicted with a~'l:ifact (62.5% precision), cognition (88.2%), state (93.3%) and t~ne (100.0%).</Paragraph>
    <Paragraph position="1"> For evaluating the rule's overall performance, there is a baseline of 69.9% and a ceiling of 84.5% accuracy. The baseline is achieved simply by classifying each clause as a state, since this is the dominant class over the supervised test cases, t However, XSimilar baselines for comparison have been used for many classification problems (Duds and Hart, 1973), e.g., part-of-I I overalll States Events acc recall prec recall prec  WordNet category of the direct object to aspectually classify have-classes, versus ceiling (C) and baseline (B) approaches.</Paragraph>
    <Paragraph position="2"> this approach classifies all event clauses incorrectly, achieving an event rr~21 of 0.0%. The ceiling of 84.5% is the maximum achievable by a rule such as ours since the first WordNet category of the direct object is not always enough to resolve aspectual ambiguity; the same category appears in both stative and event test cases.</Paragraph>
    <Paragraph position="3"> Overall classification performance using Word-Net categories is greatly improved over the baseline method. As shown in Table 7, an accuracy of 79.6% was achieved. A binomial test showed that this improvement over the baseline is significant (p &lt; .04). An event greater improvement over the baseline is illustrated by the increase in the number of event clauses correctly classified, i.e. event rrPSall. As shown in Table 7, an event recall of 67.7% was achieved by the classification rule, as compared to speech tagging (Church, 1988; Alien, 1995).</Paragraph>
    <Paragraph position="4">  the 0.0% event recall achieved by the baseline, while suffering no loss in overall accuracy. This difference in recall is more dramatic than the accuracy improvement because of the dominance of stative clauses in the test set. A favorable tradeoff in recall with no loss in accuracy presents an advantage for applications that weigh the identification of nondominant instances more heavily (Cardie and Howe, 1997). For example, it is advantageous for a medical system that identifies medical procedures to identify event clauses, since procedures are a type of event.</Paragraph>
    <Paragraph position="5"> There are several problematic cases that illustrate limitations to our approach. In particular, lexical ambiguity is mi.qleading for the task of classifying have-clauses. For example, The paticnt had Med/c~/d denotes a state, but WordNet categorizes Med/ca/d as an act. Similarly, PET, EMUand CATare categorized as animal, as shown in Table 6. This would be solved by recognizing these as proper nouns or acronyms due to capitalization. However, other ambiguous objects are more difficult to address. For e~Ample, The patient had an enema describes an event, but WordNet lists enema as artifacl; before act. As another example, The patient had a urine culture is an event, but WordNet's first sense of cu/tuw is group. Furthermore, the direct object of 10.9% of have-clauses in the medical reports are unknown to WordNet (&amp;quot;N/A&amp;quot;). This includes medical terminology, e.g., anticont~ants and vitrectomy, as well as certain expressions parsed by the English Slot Grammar that require further post-processing, such as bettoeen 39 and 29.</Paragraph>
  </Section>
  <Section position="8" start_page="13" end_page="13" type="metho">
    <SectionTitle>
7 WordNet for Linguistic Indicators
</SectionTitle>
    <Paragraph position="0"> Aspectual classification is a large-scale, domain-dependent problem. Although a complete aspectual lexicon of verbs may suffice to classify many clauses by their main verb only, a verb's primary class is often domain-dependent. For example, while many dom~inR primarily use show as an event, its appearances in medical discharge snmmaxies primarily denote states. Therefore, it is necessary to produce a specialized lexicon for each domain.</Paragraph>
    <Paragraph position="1"> One statistical approach is to measure linguistic indicators over a corpus (Siegel, 1998). These indicators measure how frequently each verb appears with markers such as those in Table 1. For example, a verb that appears more frequently in the progressive is more likely to describe an event than a state (Klavans and Chodorow, 1992). However, this approach attempts to classif T verbs independent of their context.</Paragraph>
    <Paragraph position="2"> Incorporating additional constituents of a clause could alleviate this problem. For example, indicators could be measured over verb-object pairs. However, since both the main verb and the head of the direct object are open-class categories, indicators would be sparsely measured (enjopturnips is rare).</Paragraph>
    <Paragraph position="3"> To alleviate sparsity, but retain information about  the main verb (V), indicators over verb and object's WordNet category pairs (W), and a baseline (B).</Paragraph>
    <Paragraph position="4"> the direct object, we measured indicators over verbobject-category pairs, using WordNet to derive the semantic category of each object. I describe such experiments briefly here; Further details regarding these experiments is given by Siegel (1998).</Paragraph>
    <Paragraph position="5"> Fourteen such indicators were evaluated for distinguishing clauses according to completednese over an unrestricted set of verbs and direct objects. A corpus of 75,289 parsed clauses from ten novels was used to measure indicator values. 307 training cases (196 culminated) and 308 test cases (195 culminated) were manually annotated using linguistic tests. Decision tree induction was performed over the training cases to combine the indicators.</Paragraph>
    <Paragraph position="6"> Indicators measured over the main verb and direct object category achieved a more favorable recall tradeoff than those measured over the verb only, with comparable performance in accuracy. As shown in Table 8, indicators measured over the main verb and direct object category achieved a non-culminated recall of 53.1%, as compared to 38.1% achieved by the verb-only indicators. The baseline of 63.3% accuracy is achieved by simply classifying every clause as culminated.</Paragraph>
  </Section>
class="xml-element"></Paper>