<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-2033">
  <Title>Learning to Distinguish PP Arguments from Adjuncts</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Words differ in the subcategorisation frames that realise their semantic arguments, and a given word may have several different subcategorisation frames. The subcategorisation frame includes all the complements of a given word.</Paragraph>
    <Paragraph position="1"> For instance, the sentences: +(1)John ate +(2)John ate the apple represent the intransitive and transitive frames, respectively, and both are valid frames associated with the word eat. Given that the subcategorisation frame of a given word should only include a given constituent if it is an argument, one problem is caused by the ambiguous nature of some constituents, that can be either arguments or adjuncts.</Paragraph>
    <Paragraph position="2"> The ability to distinguish between subcategorised arguments and non-subcategorised adjuncts is of great importance for several applications, such as automatic acquisition of subcategorisation lexicons from data, and this problem has been widely investigated. For instance, Buchholz (1998) investigates this task using a memory-based learning approach, where the use of syntactic and contextual features results in a 91.6% accuracy in distinguishing arguments from adjuncts. Brent (1994) looks at the problem from a more psychologically oriented perspective, trying to simulate the environment available to a human language learner, and using binomial error estimation to derive subcategorisation frames for verbs, based on imperfectly reliable local syntactic cues. This technique is able to capture the fact that the relative frequency of a verb-argument sequence is likely to be higher than that of a verb-adjunct sequence. However, the cues used in the simulations are too simple to achieve high accuracy. Steedman (1994) suggests the use of semantic information to deal with this ambiguity, given that syntax should be as close as possible to semantics. Then, given that for a particular language there is a strong correlation between the subcategorisation frames and predicate-argument structure of a given word, from the predicate-argument structure of a word it is possible to infer its subcategorisation frame.</Paragraph>
    <Paragraph position="3"> In terms of the difficulty of this task, Buchholz (1998) found that in the experiments conducted the ambiguity presented by Prepositional Phrases (PPs) was the most difficult case to classify, accounting for 23% of the errors.</Paragraph>
    <Paragraph position="4"> Moreover, Brent (1994) also found in his simulations that locative adjuncts were sometimes mistaken for arguments. In this paper we focus on the problem of distinguishing between locative PPs as arguments or adjuncts, where only if a given locative PP is an argument is that it should be included in the subcategorisation frame of the verb. The approach proposed here is to use semantically motivated preposition selection and frequency information to determine if a locative PP is an argument of the verb or if it is an adjunct. In order to test this approach, we use a computational learning system, and the results obtained indicate the effectiveness of the approach.</Paragraph>
    <Paragraph position="5"> The wider goal of this project is to investigate the process of grammatical acquisition from data. Thus, in section 2 we start by giving some background in language acquisition employed in the learning model, which is described in section 3. Characteristics of the ambiguity between arguments and adjuncts are discussed in section 4 together with the approach used to distinguish them. In section 5 we describe an experiment conducted to test the approach. We finish with some conclusions and a discussion of future work.</Paragraph>
  </Section>
class="xml-element"></Paper>