<?xml version="1.0" standalone="yes"?>
<Paper uid="J98-2002">
  <Title>Generalizing Case Frames Using a Thesaurus and the MDL Principle</Title>
  <Section position="2" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> We address the problem of automatically acquiring case frame patterns (selectional patterns, subcategorization patterns) from large corpora. A satisfactory solution to this problem would have a great impact on various tasks in natural language processing, including the structural disambiguation problem in parsing. The acquired knowledge would also be helpful for building a lexicon, as it would provide lexicographers with word usage descriptions.</Paragraph>
    <Paragraph position="1"> In our view, the problem of acquiring case frame patterns involves the following two issues: (a) acquiring patterns of individual case frame slots; and (b) learning dependencies that may exist between different slots. In this paper, we confine ourselves to the former issue, and refer the interested reader to Li and Abe (1996), which deals with the latter issue.</Paragraph>
    <Paragraph position="2"> The case frame (case slot) pattern acquisition process consists of two phases: extraction of case frame instances from corpus data, and generalization of those instances to case frame patterns. The generalization step is needed in order to represent the input case frame instances more compactly as well as to judge the (degree of) acceptability of unseen case frame instances. For the extraction problem, there have been various methods proposed to date, which are quite adequate (Hindle and Rooth 1991; Grishman and Sterling 1992; Manning 1992; Utsuro, Matsumoto, and Nagao 1992; Brent 1993; Smadja 1993; Grefenstette 1994; Briscoe and Carroll 1997). The generalization problem, in contrast, is a more challenging one and has not been solved completely. A number of methods for generalizing values of a case frame slot for a verb have been * C&amp;C Media Res. Labs., NEC Corporation, 4-1-1 Miyazaki Miyamae-ku, Kawasaki 216, Japan. E-mail: { |ihang,abe }@ccm.cl.nec.co.jp @ 1998 Association for Computational Linguistics Computational Linguistics Volume 24, Number 2 proposed. Some of these methods make use of prior knowledge in the form of an existing thesaurus (Resnik 1993a, 1993b; Framis 1994; Almuallim et al. 1994; Tanaka 1996; Utsuro and Matsumoto 1997), while others do not rely on any prior knowledge (Pereira, Tishby, and Lee 1993; Grishman and Sterling 1994; Tanaka 1994). In this paper, we propose a new generalization method, belonging to the first of these two categories, which is both theoretically well-motivated and computationally efficient.</Paragraph>
    <Paragraph position="3"> Specifically, we formalize the problem of generalizing values of a case frame slot for a given verb as that of estimating a conditional probability distribution over a partition of words, and propose a new generalization method based on the Minimum Description Length principle (MDL): a principle of data compression and statistical estimation from information theory. 1 In order to assist with efficiency, our method makes use of an existing thesaurus and restricts its attention on those partitions that are present as &amp;quot;cuts&amp;quot; in the thesaurus tree, thus reducing the generalization problem to that of estimating a &amp;quot;tree cut model&amp;quot; of the thesaurus tree. We then give an efficient algorithm that provably obtains the optimal tree cut model for the given frequency data of a case slot, in the sense of MDL. In order to test the effectiveness of our method, we conducted PP-attachment disambiguation experiments using the case frame patterns obtained by our method. Our experimental results indicate that the proposed method improves upon or is at least comparable to existing methods.</Paragraph>
    <Paragraph position="4"> The remainder of this paper is organized as follows: In Section 2, we formalize the problem of generalizing values of a case frame slot as that of estimating a conditional distribution. In Section 3, we describe our MDL-based generalization method. In Section 4, we present our experimental results. We then give some concluding remarks in Section 5.</Paragraph>
  </Section>
class="xml-element"></Paper>