<?xml version="1.0" standalone="yes"?>
<Paper uid="J02-3004">
  <Title>c(c) 2002 Association for Computational Linguistics The Disambiguation of Nominalizations</Title>
  <Section position="2" start_page="0" end_page="360" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> The automatic interpretation of compound nouns has been a long-standing problem for natural language processing (NLP). Compound nouns in English have three basic properties that present difficulties for their interpretation: (a) the compounding process is extremely productive (this means that a hypothetical system would have to interpret previously unseen instances), (b) the semantic relationship between the compound head and its modifier is implicit (this means that it cannot be easily recovered from syntactic or morphological analysis), and (c) the interpretation can be influenced by a variety of contextual and pragmatic factors.</Paragraph>
    <Paragraph position="1"> A considerable amount of effort has gone into specifying the set of semantic relations that hold between a compound head and its modifier (Levi 1978; Warren 1978; Finin 1980; Isabelle 1984). Levi (1978), for example, distinguishes two types of compound nouns: (a) compounds consisting of two nouns that are related by one of nine recoverably deletable predicates (e.g., cause relates onion tears, for relates pet spray; see the examples in (1)) and (b) nominalizations, that is, compounds whose heads are nouns derived from a verb and whose modifiers are interpreted as arguments of the related verb (e.g., a car lover loves cars; see the examples in (2)-(4)). The prenominal modifier can be either a noun or an adjective (see the examples in (2)). The nominalized verb can take a subject (see (3a)), a direct object (see (3b)) or a prepositional object  (see (3c)).</Paragraph>
    <Paragraph position="2"> (1) a. onion tears cause b. vegetable soup have [?] Division of Informatics, University of Edinburgh, 2 Buccleuch Place, Edinburgh EH8 9LW, UK. E-mail:  c. music box make d. steam iron use e. pine tree be f. night flight in g. pet spray for h. peanut butter from i. abortion problem about (2) a. parental refusal subj b. cardiac massage obj c. heart massage obj d. sound synthesizer obj (3) a. child behavior subj b. car lover obj c. soccer competition at|in (4) a. government promotion subj|obj b. satellite observation subj|obj  Besides Levi (1978), a fair number of researchers (Warren 1978; Finin 1980; Isabelle 1984; Leonard 1984) agree that there is a limited number of regularly recurring relations between a compound head and its modifier. There is far less agreement when it comes to the type and number of these relations. The relations vary from Levi's (1978) recoverably deletable predicates to Warren's (1978) paraphrases and Finin's (1980) role nominals. Leonard (1984) proposes eight relations, and Warren (1978) proposes six basic relations, whereas the number of relations proposed by Finin (1980) is potentially infinite. The attempt to restrict the semantic relations between the compound head and its modifier to a prespecified number and type has been criticized by Downing (1977), who has shown (through a series of psycholinguistic experiments) that the underlying relations can be influenced by a variety of pragmatic factors and cannot therefore be presumed to be easily enumerable. Sparck Jones (1983, page 4) further notes &amp;quot;that observations about the semantic relation holding between the compound head and its modifier can only be remarks about tendencies and not about absolutes.&amp;quot; Consider, for instance, the compound onion tears (see (1a)). The relationship cause is one of the possible interpretations the compound may receive. One could easily imagine a context in which the tears are for or about the onion. Consider example  (5a), taken from Downing (1977, page 818). Here apple-juice seat refers to the situation in which someone is instructed to sit in a seat in front of which a glass of apple juice has been placed. Given this particular state of affairs, none of the relations in (1) can be used to successfully interpret apple-juice seat. Such considerations have led Selkirk (1982) to  Lapata The Disambiguation of Nominalizations claim that only nominalizations are amenable to linguistic characterization, leaving all other compounds to be explained by pragmatics or discourse. A similar approach is put forward by Hobbs et al. (1993) for all types of compounds, including nominalizations: any two nouns can be combined, and the relation between these nouns is entirely underspecified, to be resolved pragmatically.</Paragraph>
    <Paragraph position="3">  (5) a. A friend of mine was once instructed to sit in the apple-juice seat. b. By the end of the 1920s, government promotion of agricultural development in Niger was limited, consisting mainly of crop trials and model sheep and ostrich farms.</Paragraph>
    <Paragraph position="4">  Less controversy arises with regard to nominalizations, perhaps because of the small number of allowable relations. Most approaches follow Levi (1978) in distinguishing nominalizations as a separate class of compounds, the exception being Finin (1980), who claims that most compounds are nominalizations, even in cases in which the head noun is not morphologically derived from a verb (see the examples in (1)). Under Finin's analysis the head book in the compound recipe book is a role nominal, that is, a noun that refers to a particular thematic role of another concept. This means that book refers to the object role of write, which is filled by recipe. It is not clear, however, how the implicit verb is to be recovered or why write is more appropriate than read in this example.</Paragraph>
    <Paragraph position="5"> Despite the small number of relations between the nominalized head and its modifier, the interpretation of nominalizations can readily change in different contexts. In some cases, the relation of the modifier and the nominalized verb (e.g., subject or object) can be predicted either from the subcategorization properties of the verb or from the semantics of the nominalization suffix of the head noun. Consider (3a), for example. Here child can be only the subject of behavior, since the verb behave is intransitive. In (3b) the agentive suffix -er of the head noun lover indicates that the modifier car is the object of the verb love. In other cases, the relation of the modifier and the head noun is genuinely ambiguous. Out of context the compounds government promotion and satellite observation (see example (4)) can receive either a subject or an object interpretation. One might argue that the preferred analysis for government promotion is &amp;quot;government that is promoted by someone.&amp;quot; This interpretation can be easily overridden in context, however, as shown in Example (5b): here it is the government that is doing the promotion.</Paragraph>
    <Paragraph position="6"> The automatic interpretation of compound nouns poses a challenge for empirical approaches, since the relations between a head and its modifier are not readily available in a corpus, and therefore they have to be somehow retrieved and approximated. Given the data sparseness and the parameter estimation difficulties, it is not surprising that a far greater number of symbolic than probabilistic solutions have been proposed for the automatic interpretation of compound nouns. With the exception of Wu (1993) and Lauer (1995), who use probabilistic models for compound noun interpretation (see Section 7 for details), most algorithms rely on hand-crafted knowledge bases or dictionaries that contain detailed semantic information for each noun; a sequence of rules exploit a knowledge base to choose the correct interpretation for a given compound (Finin 1980; McDonald 1982; Leonard 1984; Vanderwende 1994).</Paragraph>
    <Paragraph position="7"> In what follows we develop a probabilistic model for the interpretation of nominalizations. We focus on nominalizations whose prenominal modifier is either the underlying subject or direct object of the verb corresponding to the nominalized compound head. In other words, we focus on examples like (3a, 3b) and ignore for the moment  Computational Linguistics Volume 28, Number 3 nominalizations whose heads correspond to verbs taking prepositional complements (see example (3c)). Nominalizations are attractive from an empirical perspective: the amount of relations is small (i.e., subject or object, at least if one focuses on direct objects only) and fairly uncontroversial (see the discussion above). Although the relations are not attested in the corpus, they can be retrieved and approximated through parsing. The probabilistic interpretation of nominalizations can provide a lower bound for the difficulty of the compound interpretation task: if we cannot interpret nominalizations successfully, there is little hope for modeling more complex semantic relations stochastically (see the examples in (1)).</Paragraph>
    <Paragraph position="8"> We present a probabilistic algorithm that treats the interpretation task as a disambiguation problem. Our approach relies on the simplifying assumption that the relation of the nominalized head and its modifier noun can be approximated by the relation of the latter and the verb from which the head is derived. This approach works insofar as the verb-argument relations from which the nominalizations are derived are attested in the corpus. We show that a large number of verb-argument configurations do not occur in the corpus, something that is perhaps not surprising considering the ease with which novel compounds are created (Levi 1978). We estimate the frequencies of unseen verb-argument pairs by experimenting with three types of smoothing techniques proposed in the literature (back-off smoothing, class-based smoothing, and distance-weighted averaging) and show that their combination achieves good performance. Furthermore, we explore the contribution of context to the disambiguation task and show that performance is increased by taking contextual features into account. Our best results are achieved by combining the predictions of our probabilistic model with contextual information.</Paragraph>
    <Paragraph position="9"> The remainder of this article is organized as follows: in Section 2 we present a simple statistical model for the interpretation of nominalizations and describe the procedure used to collect the data for our experiments. Section 3 presents details on how the parameters of the model were estimated and gives a brief overview on the smoothing methods with which we experimented. Section 4 describes the algorithm used for the interpretation of nominalizations, and Section 5 reports the results of several experiments that achieve a combined accuracy of 86.1% on the British National Corpus (BNC). Section 6 discusses the findings. In Section 7 we review related work, and we conclude in Section 8.</Paragraph>
  </Section>
class="xml-element"></Paper>