<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2420">
  <Title>Two-Phase Semantic Role Labeling based on Support Vector Machines</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> A semantic role in a language is a semantic relationship between a syntactic constituent and a predicate. The shared task of CoNLL-2004 relates to recognize semantic roles in English (X. Carreras, 2004). Given a sentence, the task is to analyze a proposition expressed by a target verb of a sentence. Especially, for each target verb, all constituents in a sentence which fill semantic roles of the verb have to be recognized. This task is based only on partial parsing information, avoiding use of a full parser and external lexico-semantic knowledge base. According to previous results of the CoNLL shared task, the POS tagged, chunked, clause identified, and named-entity recognized sentences are given as an input (Figure 1).</Paragraph>
    <Paragraph position="1"> SVM is a well-known machine learning algorithm with high generalization performance in high dimensional feature spaces (H. Yamada, 2003). Also, learning with combination of multiple features is possible by virtue of polynomial kernel functions. However, since it is a binary classifier, we are often confronted with the unbalanced class distribution problem in a multiclass classification task. The larger the number of classes, the more severe the problem is. The semantic role labeling can be formulated as a multiclass classification problem. If we try to apply SVMs in the semantic role labeling problem, we have to find a method of resolving the unbalanced class distribution problem.</Paragraph>
    <Paragraph position="2"> Conceptually, semantic role labeling can be divided into two subtasks: the identification task which finds the boundary of semantic arguments in a given sentence, and the classificiation task which determines the semantic role of the argument. This provides us a hint of using SVMs with less severe unbalanced class distribution. In this paper, we present a two-phase semantic role labeling method which consists of an identification phase and a classification phase. By taking two phase model based on SVMs, we can alleviate the unbalanced class distribution problem. That is, since we find only the boundary of an argument in the identification phase, the number of classes is decreased into two (ARG, NON-ARG) or three (B-ARG, I-ARG, O). Therefore, we have to build only one or three SVM classifiers. We can alleviate the unbalanced class distribution problem by decreasing the number of negative examples, which is much larger than the number of positive exampels without two-phase modeling. In the classification phase, we classify only the identified argument into a proper semantic role. This enables us to reduce the computational cost by ignoring the non-argument constitutents.</Paragraph>
    <Paragraph position="3"> Since features for identifying arguments are different from features for classifying a role, we need to determine different feature sets appropriate for the tasks. For identification, we focus on the features to detect the dependency between a constituent and a predicate because the arguments are dependent on the predicate. For semantic role labeling, we consider both the syntactic and the semantic information such as the sentential form of the target predicate, the head of a constituent, and so on. In the following sections, we will explain the two phase semantic role labeling method in detail and show some experimental results.</Paragraph>
    <Paragraph position="4">  boundary, its named-entity tag, the target predicate, the result of semantic role labeling in the target predicate exist, and deliver.</Paragraph>
  </Section>
class="xml-element"></Paper>