<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2403">
  <Title>A Semantic Kernel for Predicate Argument Classification</Title>
  <Section position="2" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Several linguistic theories, e.g. (Jackendoff, 1990), claim that semantic information in natural language texts is connected to syntactic structures. Hence, to deal with natural language semantics, the learning algorithm should be able to represent and process structured data. The classical solution adopted for such tasks is to convert syntax structures in a flat feature representation, which is suitable for a given learning model. The main drawback is structures may not be properly represented by flat features as: (1) these latter may not be able to capture the required properties or (2) the feature designer may not know what structure properties enable the processing of semantic information.</Paragraph>
    <Paragraph position="1"> In particular, these problems arise for semantic information represented via predicate argument structures defined on syntactic parse trees. For example, Figure 1 shows the parse tree of the sentence: &amp;quot;Paul gives a lecture in Rome&amp;quot; along with the annotation of predicate arguments.</Paragraph>
    <Paragraph position="2"> A predicate may be a verb or a noun or an adjective whereas generally Arg 0 stands for agent, Arg 1 for direct object or theme or patient and ArgM may indicate locations, as in our example. A standard for predicate argument annotation is provided in the PropBank project (Kingsbury and Palmer, 2002). It has produced one million word corpus annotated with predicate-argument structures on top of the Penn Treebank 2 Wall Street Journal texts. In this way, for a large number of the Penn TreeBank parse-trees, there are available predicate annotations in a style similar to that shown in Figure 1.</Paragraph>
    <Paragraph position="3">  In PropBank only verbs are considered to be predicates whereas arguments are labeled sequentially from Arg 0 to Arg 91. In addition to these core arguments, adjunctive arguments are marked up. They include functional tags, e.g. ArgM-DIR indicates a directional, ArgM-LOC indicates a locative and ArgM-TMP stands for a temporal.</Paragraph>
    <Paragraph position="4"> An example of PropBank markup is: 1Other arguments are: Arg 2 for indirect object or benefactive or instrument or attribute or end state, Arg 3 for start point or benefactive or attribute, Arg4 for end point and so on.  an eventual 30% state in the British Company ].</Paragraph>
    <Paragraph position="5"> Automatically recognizing the boundaries and classifying the type of arguments allows Natural Language Processing systems (e.g. Information Extraction, Question Answering or Summarization) to answer questions such as &amp;quot;Who&amp;quot;, &amp;quot;When&amp;quot;, &amp;quot;What&amp;quot;, &amp;quot;Where&amp;quot;, &amp;quot;Why&amp;quot;, and so on.</Paragraph>
    <Paragraph position="6"> Given the importance of this task for Natural Language Processing applications, several machine learning approaches for argument identification and classification have been developed (Gildea and Jurasky, 2002; Surdeanu et al., 2003; Hacioglu et al., 2003; Chen and Rambow, 2003; Gildea and Hockenmaier, 2003). Their common characteristic is the adoption of feature spaces that model predicate-argument structures in a flat representation. The major problem of this choice is that there is no linguistic theory that supports the selection of syntactic features to recognize semantic structures. As a consequence, researchers are still trying to extend the basic features with other ones, e.g. (Surdeanu et al., 2003), to improve the flat feature space.</Paragraph>
    <Paragraph position="7"> Convolution kernels are a viable alternative to flat feature representation that aims to capture the structural information in term of sub-structures. The kernel functions can be used to measure similarities between two objects without explicitly evaluating the object features. That is, we do not need to understand which syntactic feature may be suited for representing semantic data. We need only to define the similarity function between two semantic structures. An example of convolution kernel on the parse-tree space is given in (Collins and Duffy, 2002).</Paragraph>
    <Paragraph position="8"> The aim was to design a novel syntactic parser by looking at the similarity between the testing parse-trees and the correct parse-trees available for training.</Paragraph>
    <Paragraph position="9"> In this paper, we define a kernel in a semantic structure space to learn the classification function of predicate arguments. The main idea is to select portions of syntactic/semantic trees that include the target &lt;predicate, argument&gt; pair and to define a kernel function between these objects. If our similarity function is well defined the learning model will converge and provide an effective argument classification.</Paragraph>
    <Paragraph position="10"> Experiments on PropBank data show not only that Support Vector Machines (SVMs) trained with the proposed semantic kernel converge but also that they have a higher accuracy than SVMs trained with a linear kernel on the standard features proposed in (Gildea and Jurasky, 2002). This provides a piece of evidence that convolution kernel can be used to learn semantic linguistic structures. Moreover, interesting research lines on the use of kernel for NLP are enabled, e.g. question classification in Question/Answering or automatic template designing in Information Extraction.</Paragraph>
    <Paragraph position="11"> The remaining of this paper is organized as follows: Section 2 defines the Predicate Argument Extraction problem and the standard solution to solve it. In Section 3 we present our approach based on the parse-tree kernel whereas in Section 4 we show our comparative results between SVMs using standard features and the proposed kernel. Finally, Section 5 summarizes the conclusions.</Paragraph>
  </Section>
class="xml-element"></Paper>