<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3213">
  <Title>Unsupervised Semantic Role Labelling</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Semantic annotation of text corpora is needed to support tasks such as information extraction and question-answering (e.g., Riloff and Schmelzenbach, 1998; Niu and Hirst, 2004). In particular, labelling the semantic roles of the arguments of a verb (or any predicate), as in (1) and (2), provides crucial information about the relations among event participants. null  1. Kivaa2a4a3a6a5a8a7a10a9a12a11a14a13a4a9a12a15a17a16a18a9a19a11a8a20 admires Matsa2a4a21a23a22a25a24a27a26a19a9a14a20 2. Joa2a29a28a31a30a25a9a12a15a27a32a33a20 returned to Londona2a4a34a35a9a14a26a18a32a36a13a4a15a17a22a8a32a36a13a36a37a38a15a17a20  Because of the importance of this task, a number of recent methods have been proposed for automatic semantic role labelling (e.g., Gildea and Jurafsky, 2002; Gildea and Palmer, 2002; Chen and Rambow, 2003; Fleischman et al., 2003; Hacioglu et al., 2003; Thompson et al., 2003). These supervised methods are limited by their reliance on the manually roletagged corpora of FrameNet (Baker et al., 1998) or PropBank (Palmer et al., 2003) as training data, which are expensive to produce, are limited in size, and may not be representative.</Paragraph>
    <Paragraph position="1"> We have developed a novel method of unsupervised semantic role labelling that avoids the need for expensive manual labelling of text, and enables the use of a large, representative corpus. To achieve this, we take a &amp;quot;bootstrapping&amp;quot; approach (e.g., Hindle and Rooth, 1993; Yarowsky, 1995; Jones et al., 1999), which initially makes only the role assignments that are unambiguous according to a verb lexicon. We then iteratively: create a probability model based on the currently annotated semantic roles, use this probability model to assign roles that are deemed to have sufficient evidence, and add the newly labelled arguments to our annotated set. As we iterate, we gradually both grow the size of the annotated set, and relax the evidence thresholds for the probability model, until all arguments have been assigned roles.</Paragraph>
    <Paragraph position="2"> To our knowledge, this is the first unsupervised semantic role labelling system applied to general semantic roles in a domain-general corpus. In a similar vein of work, Riloff and colleagues (Riloff and Schmelzenbach, 1998; Jones et al., 1999) used bootstrapping to learn &amp;quot;case frames&amp;quot; for verbs, but their approach has been applied in very narrow topic domains with topic-specific roles. In other work, Gildea (2002) has explored unsupervised methods to discover role-slot mappings for verbs, but not to apply this knowledge to label text with roles.</Paragraph>
    <Paragraph position="3"> Our approach also differs from earlier work in its novel use of classes of information in backing off to less specific role probabilities (in contrast to using simple subsets of information, as in Gildea and Jurafsky, 2002). If warranted, we base our decisions on the probability of a role given the verb, the syntactic slot (syntactic argument position), and the noun occurring in that slot. For example, the assignment to the first argument of sentence (1) above may be based on a39a41a40 Experiencera42 a43a45a44a47a46a49a48a18a50a52a51a54a53 subjecta53a8a55a56a48a18a57a45a43a59a58 . When backing off from this probability, we use statistics over more general classes of information, such as conditioning over the semantic class of the verb instead of the verb itself--for this example, psychological state verbs. Our approach yields a very simple probability model which emphasizes class-based generalizations.</Paragraph>
    <Paragraph position="4"> The first step in our algorithm is to use the verb lexicon to determine the argument slots and the roles available for them. In Section 2, we discuss the lexicon we use, and our initial steps of syntactic frame matching and &amp;quot;unambiguous&amp;quot; role assignment. This unambiguous data is leveraged by using those role assignments as the basis for the initial estimates for the probability model described in Section 3. Section 4 presents the algorithm which brings these two components together, iteratively updating the probability estimates as more and more data is labelled. In Section 5, we describe details of the materials and methods used for the experiments presented in Section 6. Our results show a large improvement over an informed baseline. This kind of unsupervised approach to role labelling is quite new, and we conclude with a discussion of limitations and on-going work in Section 7.</Paragraph>
  </Section>
class="xml-element"></Paper>