<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-0511">
  <Title>ClassifyingtheSemanticRelationsinNounCompoundsviaa Domain-SpecificLexicalHierarchy</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
2 RelatedWork
</SectionTitle>
    <Paragraph position="0"> Several approaches have been proposed for empirical noun compound interpretation. Lauer and Dras (1994) point out that there are three components to the problem: identification of the compound from within the text, syntactic analysis of the compound (left versus right association), and the interpretationoftheunderlyingsemantics. Severalresearchers have tackled the syntactic analysis (Lauer, 1995; Pustejovsky et al., 1993; Liberman and Sproat, 1992), usually using a variation of the idea of finding the subconstituents elsewhere in the corpus and using those to predict how the larger compounds are structured.</Paragraph>
    <Paragraph position="1"> We are interested in the third task, interpretation of the underlying semantics. Most related work relies on hand-written rules of one kind or another.</Paragraph>
    <Paragraph position="2"> Finin (1980) examines the problem of noun compound interpretation in detail, and constructs a complexset of rules. Vanderwende (1994) uses a sophisticated system to extract semantic information automatically from an on-line dictionary, and then manipulates a set of hand-written rules with handassigned weights to create an interpretation. Rindflesch et al. (2000) use hand-coded rule based systems to extract the factual assertions from biomedical text. Lapata (2000) classifies nominalizations according to whether the modifier is the subject or the object of the underlying verb expressed by the head noun.</Paragraph>
    <Paragraph position="3">  In the related sub-area of information extraction (Cardie, 1997; Riloff, 1996), the main goal is to find every instance of particular entities or events of interest. These systems use empirical techniques to learn which terms signal entities of interest, in order to fill in pre-defined templates. Our goals are more general than those of information extraction, and so this work should be helpful for that task. However, our approach will not solve issues surrounding previously unseen proper nouns, which are often important for information extraction tasks.</Paragraph>
    <Paragraph position="4"> There have been several efforts to incorporate lexical hierarchies into statistical processing, primarily for the problem of prepositional phrase (PP) attachment. The current standard formulation is: given a verb followed by a noun and a prepositional phrase, represented by the tuple v,n1,p,n2, determine which of v or n1 the PP consisting of p and n2 attaches to, or is most closely associated with.</Paragraph>
    <Paragraph position="5"> Because the data is sparse, empirical methods that train on word occurrences alone (Hindle and Rooth, 1993) have been supplanted by algorithms that generalize one or both of the nouns according to classmembership measures (Resnik, 1993; Resnik and Hearst, 1993; Brill and Resnik, 1994; Li and Abe, 1998), but the statistics are computed for the particular preposition and verb.</Paragraph>
    <Paragraph position="6"> It is not clear how to use the results of such analysis after they are found; the semantics of the rela- null Nominalizations are compounds whose head noun is a nominalized verb and whose modifier is either the subject or the object of the verb. We do not distinguish the NCs on the basis of their formation.</Paragraph>
    <Paragraph position="7"> tionshipbetweenthetermsmuststillbedetermined.</Paragraph>
    <Paragraph position="8"> In our framework we would cast this problem as findingtherelationship R(p,n2)thatbestcharacterizes the preposition and the NP that follows it, and then seeing if the categorization algorithm determines their exists any relationship R</Paragraph>
    <Paragraph position="10"> fact that they condition probabilities on a particular verb and noun. Resnik (1993; 1995) use classes in Wordnet (Fellbaum, 1998) and a measure of conceptual association to generalize over the nouns. Brill and Resnik (1994) use Brill's transformation-based algorithm along with simple counts within a lexical hierarchy in order to generalize over individual words. Li and Abe (1998) use a minimum description length-based algorithm to find an optimal tree cut over WordNet for each classification problem, finding improvements over both lexical association (Hindle and Rooth, 1993) and conceptual association, and equaling the transformation-based results. Our approach differs from these in that we are using machine learning techniques to determine which level of the lexical hierarchy is appropriate for generalizing across nouns.</Paragraph>
  </Section>
class="xml-element"></Paper>