<?xml version="1.0" standalone="yes"?>
<Paper uid="J04-1003">
  <Title>c(c) 2004 Association for Computational Linguistics Verb Class Disambiguation Using Informative Priors</Title>
  <Section position="2" start_page="0" end_page="50" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Much research in lexical semantics has concentrated on the relation between verbs and their arguments. Many scholars hypothesize that the behavior of a verb, particularly with respect to the expression and interpretation of its arguments, is to a large extent determined by its meaning (Talmy 1985; Jackendoff 1983; Goldberg 1995; Levin 1993; Pinker 1989; Green 1974; Gropen et al. 1989; Fillmore 1965). The correspondence between verbal meaning and syntax has been extensively studied in Levin (1993), which argues that verbs which display the same diathesis alternations--alternations in the realization of their argument structure--can be assumed to share certain meaning components and to form a semantically coherent class.</Paragraph>
    <Paragraph position="1"> The converse of this assumption is that verb behavior (i.e., participation in diathesis alternations) can be used to provide clues about aspects of meaning, which in turn can be exploited to characterize verb senses (referred to as classes in Levin's [1993] terminology). A major advantage of this approach is that criteria for assigning senses can be more concrete than is traditionally assumed in lexicographic work (e.g., WordNet or machine-readable dictionaries) concerned with sense distinctions (Palmer 2000). As an example consider sentences (1)-(4), taken from Levin. Examples (1) and (2) illustrate the dative and benefactive alternations, respectively. Dative verbs alternate between the prepositional frame &amp;quot;NP1 V NP2 to NP3&amp;quot; (see (1a)) and the double-object frame &amp;quot;NP1 V NP2 NP3&amp;quot; (see (1b)), whereas benefactive verbs alternate between the double-object frame (see (2a)) and the prepositional frame &amp;quot;NP1 V NP2 for NP3&amp;quot; (see (2b)). To decide whether a verb is benefactive or dative it suffices to test the acceptability of the for and to frames. Verbs undergoing the conative alternation can be attested either as transitive or as intransitive with a prepositional phrase headed by the word at.</Paragraph>
    <Paragraph position="2">  The role filled by the object of the transitive variant is shared by the noun phrase complement of at in the intransitive variant (see (3)). This example makes explicit that class assignment depends not only on syntactic facts but also on judgments about [?] Department of Computer Science, Regent Court, 211 Portobello Street, Sheffield, S1 4DP, UK. E-mail: mlap@dcs.shef.ac.uk.</Paragraph>
    <Paragraph position="3"> + Department of Linguistics, Oxley Hall,1712 Neil Avenue, Columbus, OH. E-mail: cbrew@ling.ohiostate.edu. null  Computational Linguistics Volume 30, Number 1 semantic roles. Similarly, the possessor object alternation involves a possessor and a possessed attribute that can be manifested either as the verbal object or as the object of a prepositional phrase headed by for (see (4)).</Paragraph>
    <Paragraph position="4">  (1) a. Bill sold a car to Tom.</Paragraph>
    <Paragraph position="5"> b. Bill sold Tom a car.</Paragraph>
    <Paragraph position="6"> (2) a. Martha carved the baby a toy.</Paragraph>
    <Paragraph position="7"> b. Martha carved a toy for the baby.</Paragraph>
    <Paragraph position="8"> (3) a. Paula hit the fence.</Paragraph>
    <Paragraph position="9"> b. Paula hit at the fence.</Paragraph>
    <Paragraph position="10"> (4) a. I admired his honesty.</Paragraph>
    <Paragraph position="11">  b. I admired him for his honesty.</Paragraph>
    <Paragraph position="12"> Observation of the semantic and syntactic behavior of pay and give reveals that they pattern with sell in licensing the dative alternation. These verbs are all members of the Give class. Verbs like make and build behave similarly to carve in licensing the benefactive alternation and are members of the class of Build verbs. The verbs beat, kick, and hit undergo the conative alternation; they are all members of the Hit verb class. By grouping together verbs that pattern together with respect to diathesis alternations, Levin (1993) defines approximately 200 verb classes, which she argues reflect important semantic regularities. These analyses (and many similar ones by Levin and her successors) rely primarily on straightforward syntactic and syntactico-semantic criteria. To adopt this approach is to accept some limitations on the reach of our analyses, since not all semantically interesting differences will have the appropriate reflexes in syntax. Nevertheless, the emphasis on concretely available observables makes Levin's methodology a good candidate for automation (Palmer 2000).</Paragraph>
    <Paragraph position="13"> Therefore, Levin's (1993) classification has formed the basis for many efforts that aim to acquire lexical semantic information from corpora. These exploit syntactic cues, or at least cues that are plausibly related to syntax (Merlo and Stevenson 2001; Schulte im Walde 2000; Lapata 1999; McCarthy 2000). Other work has used Levin's classification (in conjunction with other lexical resources) to create dictionaries that express the systematic correspondence between syntax and meaning (Dorr 1997; Dang, Rosenzweig, and Palmer 1997; Dorr and Jones 1996). Levin's inventory of verbs and classes has been also useful for applications such as machine translation (Dorr 1997; Palmer and Wu 1995), generation (Stede 1998), information retrieval (Levow, Dorr, and Lin 2000), and document classification (Klavans and Kan 1998).</Paragraph>
    <Paragraph position="14"> Although the classification provides a general framework for describing verbal meaning, it says only which verb meanings are possible, staying silent on the relative likelihoods of the different meanings. The inventory captures systematic regularities in the meaning of words and phrases but falls short of providing a probabilistic model of these regularities. Such a model would be useful in applications that need to resolve ambiguity in the presence of multiple and conflicting probabilistic constraints. More precisely, Levin (1993) provides an index of 3,024 verbs for which she lists the semantic classes and diathesis alternations. The mapping between verbs and classes is not one to one. Of the 3,024 verbs which she covers, 784 are listed as having more than one class. Even though Levin's monosemous verbs outnumber her polysemous verbs by a factor of nearly four to one, the total frequency of the former (4,252,715)  Lapata and Brew Verb Class Disambiguation Using Informative Priors Table 1 Polysemous verbs according to Levin.</Paragraph>
    <Paragraph position="15">  Relation between number of classes and alternations.</Paragraph>
    <Paragraph position="16"> is comparable to the total frequency of the latter (3,986,014). This means that close to half of the cases processed by a semantic tagger would manifest some degree of ambiguity. The frequencies are detailed in Table 1 and were compiled from a lemmatized version of the British National Corpus (BNC) (Burnard 1995). Furthermore, as shown in Figure 1, the level of ambiguity increases in tandem with the number of alternations licensed by a given verb. Consider, for example, verbs participating in one alternation only: Of these, 90.4% have one semantic class, 8.6% have two classes, 0.7% have three classes, and 0.3% have four classes. In contrast, of the verbs licensing six different alternations, 14% have one class, 17% have two classes, 12.4% have three classes, 53.6% have four classes, 2% have six classes, and 1% have seven classes. As ambiguity increases, so does the availability and potential utility of information about diathesis alternations.</Paragraph>
    <Paragraph position="17"> Palmer (2000) and Dang et al. (1998) argue that syntactic frames and verb classes are useful for developing principled classifications of verbs. We go beyond this, showing that they can also be of assistance in disambiguation. Consider, for instance, the verb serve, which is a member of four Levin classes: Give, Fit, Masquerade, and Fulfilling. Each of these classes can in turn license four distinct syntactic frames.  Computational Linguistics Volume 30, Number 1 As shown in the examples  below, in (5a) serve appears ditransitively and belongs to the semantic class of Give verbs, in (5b) it occurs transitively and is a member of the class of Fit verbs, in (5c) it takes the predicative complement as minister of the interior and is a member of the class of Masquerade verbs. Finally, in sentence (5d) serve is a Fulfilling verb and takes two complements, a noun phrase (an apprenticeship) and a prepositional phrase headed by to (to a still-life photographer). In the case of verbs like serve, we can guess their semantic class solely on the basis of the frame with which  they appear.</Paragraph>
    <Paragraph position="18"> (5) a. I'm desperately trying to find a venue for the reception which can serve our guests an authentic Italian meal. NP1 V NP2 NP3 b. The airline serves 164 destinations in over 75 countries. NP1 V NP2 c. Jean-Antoine Chaptal was a brilliant chemist and technocrat who  served Napoleon as minister of the interior from 1800 to 1805. NP1 V NP2 as NP3 d. Before her brief exposure to pop stardom, she served an apprenticeship to a still-life photographer. NP1 V NP2 to NP3 But sometimes we do not have the syntactic information that would provide cues for semantic disambiguation. Consider example (6). The verb write is a member of three Levin classes, two of which (Message Transfer, Performance) take the double-object frame. In this case, we have the choice between theMessageTransferreading (see (6a)) and the Performance reading (see (6b)). The same situation arises with the verb toast, which is listed as a Prepare verb and a Judgment verb; both these classes license the prepositional frame &amp;quot;NP1 V NP2 for NP3.&amp;quot; In sentence (7a) the preferred reading is that of Prepare rather than that of Judgment (see sentence (7b)). The verb study is ambiguous among three classes when attested in the transitive frame: Learn (see example (8a)), Sight (see example (8b)), and Assessment (see example (8c)). The verb convey, when attested in the prepositional frame &amp;quot;NP1 V NP2 to NP3,&amp;quot; can be ambiguous between the Say class (see example (9a)) and the Send class (see example (9b)). In order to correctly decide the semantic class for a given ambiguous verb, we would need not only detailed semantic information about the verb's arguments, but also a considerable amount of world knowledge. Admittedly, selectional restrictions are sufficient for distinguishing (7a) from (7b) (one normally heats up inanimate entities and salutes animate ones), but selectional restrictions alone are probably not enough to disambiguate (6a) from (6b), since both letter and screenplay are likely to be described as written material. Rather, we need fine-grained world knowledge: Both scripts and letters can be written for someone: only letters can be written to someone.  (6) a. A solicitor wrote him a letter at the airport.</Paragraph>
    <Paragraph position="19"> b. I want you to write me a screenplay called &amp;quot;The Trip.&amp;quot; (7) a. He sat by the fire and toasted a piece of bread for himself.</Paragraph>
    <Paragraph position="20"> b. We all toasted Nigel for his recovery.</Paragraph>
    <Paragraph position="21"> 2 Unless otherwise stated, our example sentences were taken (possibly in simplified form) from the BNC.  Lapata and Brew Verb Class Disambiguation Using Informative Priors (8) a. Chapman studied medicine at Cambridge.</Paragraph>
    <Paragraph position="22"> b. Romanov studied the old man carefully, looking for some sign that he knew exactly what had been awaiting him at the bank.</Paragraph>
    <Paragraph position="23"> c. The alliance will also study the possibility of providing service to other high-volume products, such as IBM and multi-vendor workstations.</Paragraph>
    <Paragraph position="24"> (9) a. By conveying the news to her sister, she would convey by implication  something of her own anxiety.</Paragraph>
    <Paragraph position="25"> b. The judge signed the committal warrant and the police conveyed Mr.</Paragraph>
    <Paragraph position="26"> Butler to prison, giving the warrant to the governor.</Paragraph>
    <Paragraph position="27"> This need for world knowledge (or at least a convenient way of approximating this knowledge) is not an isolated phenomenon but manifests itself across a variety of classes and frames (e.g., double object, transitive, prepositional frame: see examples (6)-(9)). We have argued that the concreteness of Levin-style verb classes is an advantage, but this advantage would be compromised if we tried to fold too much world knowledge into the classification. We do not do this. Instead, Section 5 of the current article describes disambiguation experiments in which our probabilistic Levin classes are used in tandem with proxies for appropriate world knowledge.</Paragraph>
    <Paragraph position="28"> It is important to point out that Levin's (1993) classification is not intended as an exhaustive description of English verbs, their meanings, and their likelihood. Many other classifications could have been built using the same principles. A different grouping might, for example, have occurred if finer or coarser semantic distinctions were taken into account (see Merlo and Stevenson [2001] and Dang, Rosenzweig, and Palmer [1997] for alternative classifications) or if the containment of ambiguity was one of the classification objectives. As pointed out by Kipper, Dang, and Palmer (2000), Levin classes exhibit inconsistencies, and verbs are listed in multiple classes, some of which have conflicting sets of syntactic frames. This means that some ambiguities may also arise as a result of accidental errors or inconsistencies. The classification was created not with computational uses in mind, but for human readers, so it has not been necessary to remedy all the errors and omissions that might cause trouble for machines. Similar issues arise in almost all efforts to make use of preexisting lexical resources for computational purposes (Briscoe and Carroll 1997), so none of the above comments should be taken as criticisms of Levin's achievement.</Paragraph>
    <Paragraph position="29"> The objective of this article is to show how to train and use a probabilistic version of Levin's classification in verb sense disambiguation. We treat errors and inconsistencies in the classification as noise. Although all our tests have used Levin's classes and the British National Corpus, the method itself depends neither on the details of Levin's classification nor on parochial facts about the English language. Our future work will include tests on other languages, other classifications, and other corpora. The model developed in this article takes as input a partially parsed corpus and generates, for each combination of a verb and its syntactic frame, a probability distribution over the available verb classes. The corpus itself does not have to be labeled with classes. This makes it feasible to use large corpora. Our model is not immediately useful for disambiguation, because it cannot discriminate among different occurrences of the same verb and frame, but it can (as we show in Section 5) be used as a prior in a full disambiguation system that does take appropriate account of context. The model relies on several gross simplifications; it does not take selectional restrictions, discourse, or pragmatic information into account but is demonstrably superior to simpler priors that make no use of subcategorization.</Paragraph>
    <Paragraph position="30">  Computational Linguistics Volume 30, Number 1 The remainder of this article is organized as follows. In Section 2 we describe the probabilistic model and the estimation of the various model parameters. In Sections 3 and 4 we report on the results of two experiments that use the model to derive the dominant class for polysemous verbs. Sections 5 and 6 discuss our verb class disambiguation experiments. We base our results on the BNC, a 100-million-word collection of samples of written and spoken language from a wide range of sources designed to represent a wide cross-section of current British English, both spoken and written (Burnard 1995). We discuss our results in Section 7 and review related work in Section 8.</Paragraph>
  </Section>
class="xml-element"></Paper>