<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0638">
  <Title>Exploiting Full Parsing Information to Label Semantic Roles Using an Ensemble of ME and SVM via Integer Linear Programming</Title>
  <Section position="4" start_page="0" end_page="234" type="metho">
    <SectionTitle>
2 System Architecture
</SectionTitle>
    <Paragraph position="0"> Our SRL system is comprised of four stages: pruning, argument classification, classification model incorporation, and integer linear programming.</Paragraph>
    <Paragraph position="1"> This section describes how we build these stages, including the features used in training the argument classification models.</Paragraph>
    <Section position="1" start_page="0" end_page="233" type="sub_section">
      <SectionTitle>
2.1 Pruning
</SectionTitle>
      <Paragraph position="0"> When the full parsing tree of a sentence is available, only the constituents in the tree are considered as argument candidates. In CoNLL-2005, full parsing trees are provided by two full parsers: the Collins parser (Collins, 1999) and the Charniak parser (Charniak, 2000). According to Punyakanok et al. (2005), the boundary agreement of Charniak is higher than that of Collins; therefore, we choose the Charniak parser's results. However, there are two million nodes on the full parsing trees in the training corpus, which makes the training time of machine learning algorithms extremely long. Besides, noisy information from unrelated parts of a sentence could also affect the training of machine learning models. Therefore, our system exploits the heuristic rules introduced by Xue and Palmer (2004) to filter out simple constituents that are unlikely to be arguments. Applying pruning heuristics to the output of Charniak's parser effectively eliminates 61% of the training data and 61.3% of the development data, while still achieves 93% and 85.5% coverage of the correct arguments in the training and development sets, respectively.</Paragraph>
    </Section>
    <Section position="2" start_page="233" end_page="233" type="sub_section">
      <SectionTitle>
2.2 Argument Classification
</SectionTitle>
      <Paragraph position="0"> This stage assigns the final labels to the candidates derived in Section 2.1. A multi-class classifier is trained to classify the types of the arguments supplied by the pruning stage. In addition, to reduce the number of excess candidates mistakenly output by the previous stage, these candidates can be labeled as null (meaning &amp;quot;not an argument&amp;quot;). The  parsing tree from the parse constituent being classified to the predicate.</Paragraph>
      <Paragraph position="1"> * Constituent Type * Position - Whether the phrase is located before or after the predicate.</Paragraph>
      <Paragraph position="2"> * Voice - passive: if the predicate has a POS tag VBN, and its chunk is not a VP, or it is preceded by a form of &amp;quot;to be&amp;quot; or &amp;quot;to get&amp;quot; within its chunk; otherwise, it is active.</Paragraph>
      <Paragraph position="3"> * Head Word - calculated using the head word table described by Collins (1999).</Paragraph>
      <Paragraph position="4"> * Head POS - The POS of the Head Word.</Paragraph>
      <Paragraph position="5"> * Sub-categorization - The phrase structure rule that expands the predicate's parent node in the parsing tree.</Paragraph>
      <Paragraph position="6">  ceding and the two words following the target phrase, as well as their corresponding POSs.</Paragraph>
      <Paragraph position="7"> * Context Chunk Type - The two chunks preceding and the two chunks following the target phrase.</Paragraph>
    </Section>
    <Section position="3" start_page="233" end_page="234" type="sub_section">
      <SectionTitle>
Full Parsing Features
</SectionTitle>
      <Paragraph position="0"> We believe that information from related constituents in the full parsing tree helps in labeling the target constituent. Denote the target constituent by t. The following features are the most common baseline features of t's parent and sibling constituents. For example, Parent/ Left Sibling/ Right Sibling Path denotes t's parents', left sibling's, and right sibling's Path features.</Paragraph>
      <Paragraph position="1">  then the head of this PP is also used as a feature.</Paragraph>
      <Paragraph position="2">  We use all the features of the SVM-based and ME-based argument classification models. All SVM classifiers are realized using SVM-Light with a polynomial kernel of degree 2. The ME-based model is implemented based on Zhang's MaxEnt</Paragraph>
    </Section>
    <Section position="4" start_page="234" end_page="234" type="sub_section">
      <SectionTitle>
2.3 Classification Model Incorporation
</SectionTitle>
      <Paragraph position="0"> We now explain how we incorporate the SVM-based and ME-based argument classification models. After argument classification, we acquire two</Paragraph>
    </Section>
    <Section position="5" start_page="234" end_page="234" type="sub_section">
      <SectionTitle>
2.4 Integer Linear Programming (ILP)
</SectionTitle>
      <Paragraph position="0"> To represent full parsing information as features, there are still several syntactic constraints on a parsing tree in the SRL problem. For example, on a path of the parsing tree, there can be only one constituent annotated as a non-null argument. However, it is difficult to encode this constraint in the argument classification models. Therefore, we apply integer linear programming to resolve inconsistencies produced in the argument classification stage.</Paragraph>
      <Paragraph position="1"> According to Punyakanok et al. (2004), given a set of constituents, S, and a set of semantic role labels, A, the SRL problem can be formulated as an ILP as follows: Let z ia be the indicator variable that represents whether or not an argument, a, is assigned to any</Paragraph>
      <Paragraph position="3"> is calculated by the argument classification models. The goal of this ILP is to find a set of assignments for all z  the filter function into linear equalities or inequalities, and use them in this ILP.</Paragraph>
      <Paragraph position="5"> on the same path in a full parsing tree, only one argument can be assigned to an argument type. Thus, at least k [?] 1 arguments will be null, which is represented by ph in the following linear equality:</Paragraph>
      <Paragraph position="7"> Constraint II: No duplicate argument classes Within the same sentence, A0-A5 cannot appear more than once. The inequality for A0 is therefore:  For each verb, we look up its allowed roles. This constraint is represented by summing all the corresponding indicator variables to 0.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="234" end_page="235" type="metho">
    <SectionTitle>
3 Experiment Results
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="234" end_page="234" type="sub_section">
      <SectionTitle>
3.1 Data and Evaluation Metrics
</SectionTitle>
      <Paragraph position="0"> The data, which is part of the PropBank corpus, consists of sections from the Wall Street Journal part of the Penn Treebank. All experiments were carried out using Section 2 to Section 21 for training, Section 24 for development, and Section 23 for testing. Unlike CoNLL-2004, part of the Brown corpus is also included in the test set.</Paragraph>
    </Section>
    <Section position="2" start_page="234" end_page="235" type="sub_section">
      <SectionTitle>
3.2 Results
</SectionTitle>
      <Paragraph position="0"> Table 1 shows that our system makes little difference to the development set and Test WSJ. However, due to the intrinsic difference between the WSJ and Brown corpora, our system performs better on Test WSJ than on Test Brown.</Paragraph>
      <Paragraph position="1">  opment set.</Paragraph>
      <Paragraph position="2"> From Table 2, we can see that the model with full parsing features outperforms the model without the features in all three performance matrices. After applying ILP, the performance is improved further. We also observe that SVM slightly outperforms ME. However, the hybrid argument classification model achieves the best results in all three metrics.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>