<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0636">
  <Title>Joint Parsing and Semantic Role Labeling</Title>
  <Section position="4" start_page="0" end_page="225" type="metho">
    <SectionTitle>
2 Base SRL System
</SectionTitle>
    <Paragraph position="0"> Our approach to joint parsing and SRL begins with a base SRL system, which uses a standard architecture from the literature. Our base SRL system is a cascade of maximum-entropy classifiers which select the semantic argument label for each constituent of a full parse tree. As in other systems, we use three stages: pruning, identification, and classification. First, in pruning, we use a deterministic pre-processing procedure introduced by Xue and Palmer (2004) to prune many constituents which are almost certainly not arguments. Second, in identification, a binary MaxEnt classifier is used to prune remaining constituents which are predicted to be null with  fier.</Paragraph>
    <Paragraph position="1"> high probability. Finally, in classification, a multi-class MaxEnt classifier is used to predict the argument type of the remaining constituents. This classifer also has the option to output NULL.</Paragraph>
    <Paragraph position="2"> It can happen that the returned semantic arguments overlap, because the local classifiers take no global constraints into account. This is undesirable, because no overlaps occur in the gold semantic annotations. We resolve overlaps using a simple recursive algorithm. For each parent node that overlaps with one of its descendents, we check which predicted probability is greater: that the parent has its locally-predicted argument label and all its descendants are null, or that the descendants have their optimal labeling, and the parent is null. This algorithm returns the non-overlapping assignment with globally highest confidence. Overlaps are uncommon, however; they occurred only 68 times on the 1346 sentences in the development set.</Paragraph>
    <Paragraph position="3"> We train the classifiers on PropBank sections 0221. If a true semantic argument fails to match any bracketing in the parse tree, then it is ignored.</Paragraph>
    <Paragraph position="4"> Both the identification and classification models are trained using gold parse trees. All of our features are standard features for this task that have been used in previous work, and are listed in Tables 1 and 2.</Paragraph>
    <Paragraph position="5"> We use the maximum-entropy implementation in the Mallet toolkit (McCallum, 2002) with a Gaussian prior on parameters.</Paragraph>
  </Section>
  <Section position="5" start_page="225" end_page="225" type="metho">
    <SectionTitle>
3 Reranking Parse Trees Using SRL
</SectionTitle>
    <Paragraph position="0"> Information Here we give the general framework for the reranking methods that we present in the next section. We write a joint probability model over semantic frames F and parse trees t given a sentence x as</Paragraph>
    <Paragraph position="2"> where p(t|x) is given by a standard probabilistic parsing model, and p(F|t,x) is given by the base-line SRL model described previously.</Paragraph>
    <Paragraph position="3">  opment set by the type of parse trees used. In this paper, we choose (F[?],t[?]) to approximately maximize the probability p(F,t|x) using a reranking approach. To do the reranking, we generate a list of k-best parse trees for a sentence, and for each predicted tree, we predict the best frame using the base SRL model. This results in a list {(Fi,ti)} of parse tree / SRL frame pairs, from which the reranker chooses. Thus, our different reranking methods vary only in which parse tree is selected; given a parse tree, the frame is always chosen using the best prediction from the base model.</Paragraph>
    <Paragraph position="4"> The k-best list of parses is generated using Dan Bikel's (2004) implementation of Michael Collins' parsing model. The parser is trained on sections 221 of the WSJ Treebank, which does not overlap with the development or test sets. The k-best list is generated in Bikel's implementation by essentially turning off dynamic programming and doing very aggressive beam search. We gather a maximum of 500 best parses, but the limit is not usually reached using feasible beam widths. The mean number of parses per sentence is 176.</Paragraph>
  </Section>
class="xml-element"></Paper>