<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2607">
  <Title>Tree Kernel Engineering in Semantic Role Labeling Systems</Title>
  <Section position="2" start_page="0" end_page="49" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> A lot of attention has been recently devoted to the design of systems for the automatic labeling of semantic roles (SRL) as defined in two important projects: FrameNet (Johnson and Fillmore, 2000), inspired by Frame Semantics, and PropBank (Kingsbury and Palmer, 2002) based on Levin's verb classes. In general, given a sentence in natural language, the annotation of a predicate's semantic roles requires (1) the detection of the target word that embodies the predicate and (2) the detection and classification of the word sequences constituting the predicate's arguments. In particular, step (2) can be divided into two different phases: (a) boundary detection, in which the words of the sequence are detected and (b) argument classification, in which the type of the argument is selected.</Paragraph>
    <Paragraph position="1"> Most machine learning models adopted for the SRL task have shown that (shallow or deep) syntactic information is necessary to achieve a good labeling accuracy. This research brings a wide empirical evidence in favor of the linking theories between semantics and syntax, e.g. (Jackendoff, 1990). However, as no theory provides a sound and complete treatment of such issue, the choice and design of syntactic features for the automatic learning of semantic structures requires remarkable research efforts and intuition.</Paragraph>
    <Paragraph position="2"> For example, the earlier studies concerning linguistic features suitable for semantic role labeling were carried out in (Gildea and Jurasfky, 2002).</Paragraph>
    <Paragraph position="3"> Sincethen, researchershaveproposeddiversesyntactic feature sets that only slightly enhance the previous ones, e.g. (Xue and Palmer, 2004) or (Carreras and M`arquez, 2005). A careful analysis of such features reveals that most of them are syntactic tree fragments of training sentences, thus a natural way to represent them is the adoption of tree kernels as described in (Moschitti, 2004). The idea is to associate with each argument the minimal subtree that includes the target predicate with one of its arguments, and to use a tree kernel function to evaluate the number of common substructures between two such trees. Such approach is in linewithcurrentresearchontheuseoftreekernels for natural language learning, e.g. syntactic parsing re-ranking (Collins and Duffy, 2002), relation extraction (Zelenko et al., 2003) and named entity recognition (Cumby and Roth, 2003; Culotta and Sorensen, 2004).</Paragraph>
    <Paragraph position="4"> Regarding the use of tree kernels for SRL, in (Moschitti, 2004) two main drawbacks have been  pointed out: * Highly accurate boundary detection cannot be carried out by a tree kernel model since correct and incorrect arguments may share a large portion of the encoding trees, i.e. they may share many substructures.</Paragraph>
    <Paragraph position="5"> * Manually derived features (extended with a  polynomialkernel)havebeenshowntobesuperior to tree kernel approaches. Nevertheless, we believe that modeling a completelykernelizedSRLsystemisusefulforthefol- null lowing reasons: * We can implement it very quickly as the feature extractor module only requires the writing of the subtree extraction procedure. Traditional SRL systems are, in contrast, based on the extraction of more than thirty features (Pradhanetal., 2005), whichrequirethewriting of at least thirty different procedures. * Combining it with a traditional attribute-value SRL system allows us to obtain a more accurate system. Usually the combination of two traditional systems (based on the same machine learning model) does not result in an improvement as their features are more or less equivalent as shown in (Carreras and M`arquez, 2005).</Paragraph>
    <Paragraph position="6"> * The study of the effective structural features can inspire the design of novel linear features which can be used with a more efficient model (i.e. linear SVMs).</Paragraph>
    <Paragraph position="7"> In this paper, we carry out tree kernel engineering (Moschitti et al., 2005) to increase both accuracy and speed of the boundary detection and argument classification phases. The engineering approach relates to marking the nodes of the encoding subtrees in order to generate substructures more strictly correlated with a particular argument, boundary or predicate. For example, marking the node that exactly covers the target argument helps tree kernels to generate different substructures for correct and incorrect argument boundaries.</Paragraph>
    <Paragraph position="8"> The other technique that we applied to engineer different kernels is the subdivision of internal and pre-terminal nodes. We show that designing different classifiers for these two different node types slightly increases the accuracy and remarkably decreases the learning and classification time.</Paragraph>
    <Paragraph position="9"> An extensive experimentation of our tree kernels with Support Vector Machines on the CoNLL 2005 data set provides interesting insights on the design of performant SRL systems entirely based on tree kernels.</Paragraph>
    <Paragraph position="10"> In the remainder of this paper, Section 2 introduces basic notions on SRL systems and tree kernels. Section 3 illustrates our new kernels for both boundary and classification tasks. Section 4 shows the experiments of SVMs with the above tree kernel based classifiers.</Paragraph>
  </Section>
class="xml-element"></Paper>