<?xml version="1.0" standalone="yes"?>
<Paper uid="E99-1026">
  <Title>Japanese Dependency Structure Analysis Based on Maximum Entropy Models</Title>
  <Section position="2" start_page="0" end_page="196" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Dependency structure analysis is one of the basic techniques in Japanese sentence analysis. The Japanese dependency structure is usually represented by the relationship between phrasal units called 'bunsetsu.' The analysis has two conceptual steps. In the first step, a dependency matrix is prepared. Each element of the matrix represents how likely one bunsetsu is to depend on the other. In the second step, an optimal set of dependencies for the entire sentence is found. In this paper, we will mainly discuss the first step, a model for estimating dependency likelihood.</Paragraph>
    <Paragraph position="1"> So far there have been two different approaches to estimating the dependency likelihood, One is the rule-based approach, in which the rules are created by experts and likelihoods are calculated by some means, including semiautomatic corpus-based methods but also by manual assignment of scores for rules. However, hand-crafted rules have the following problems.</Paragraph>
    <Paragraph position="2"> * They have a problem with their coverage. Because there are many features to find correct dependencies, it is difficult to find them manually. null * They also have a problem with their consistency, since many of the features compete with each other and humans cannot create consistent rules or assign consistent scores.</Paragraph>
    <Paragraph position="3"> * As syntactic characteristics differ across different domains, the rules have to be changed when the target domain changes. It is costly to create a new hand-made rule for each domain. null At/other approach is a fully automatic corpus-based approach. This approach has the potential to overcome the problems of the rule-based approach. It automatically learns the likelihoods of dependencies from a tagged corpus and calculates the best dependencies for an input sentence. We take this approach. This approach is taken by some other systems (Collins, 1996; Fujio and Matsumoto, 1998; Haruno et ah, 1998). The parser proposed by Ratnaparkhi (Ratnaparkhi, 1997) is considered to be one of the most accurate parsers in English. Its probability estimation is based on the maximum entropy models. We also use the maximum entropy model. This model learns the weights of given features from a training corpus.</Paragraph>
    <Paragraph position="4"> The weights are calculated based on the frequencies of the features in the training data. The set of features is defined by a human. In our model, we use features of bunsetsu, such as character strings, parts of speech, and inflection types of bunsetsu, as well as information between bunsetsus, such as the existence of punctuation, and the distance between bunsetsus. The probabilities of dependencies are estimated from the model by using those features in input sentences. We assume that the overall dependencies in a whole sentence can be determined as the product of the probabilities of all the dependencies in the sentence.</Paragraph>
    <Paragraph position="5">  Now, we briefly describe the algorithm of dependency analysis. It is said that Japanese dependencies have the following characteristics.</Paragraph>
    <Paragraph position="6">  (1) Dependencies are directed from left to right (2) Dependencies do not cross (3) A bunsetsu, except for the rightmost one, depends on only one bunsetsu (4) In many cases, the left context is not neces null sary to determine a dependency 1 The analysis method proposed in this paper is designed to utilize these features. Based on these properties, we detect the dependencies in a sentence by analyzing it backwards (from right to left). In the past, such a backward algorithm has been used with rule-based parsers (e.g., (Fujita, 1988)). We applied it to our statistically based approach. Because of the statistical property, we can incorporate a beam search, an effective way of limiting the search space in a backward analysis.</Paragraph>
  </Section>
class="xml-element"></Paper>