<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1622">
  <Title>Semantic Role Labeling via Instance-Based Learning</Title>
  <Section position="4" start_page="0" end_page="180" type="intro">
    <SectionTitle>
2 Related Work
Features
</SectionTitle>
    <Paragraph position="0"> Most of the systems outlined in CoNLL2004 and CoNLL2005 utilize as many as 30 features for learning approaches to SRL. The research presented here uses only seven of these:  from a constituent &amp;quot;The officer&amp;quot; to the predicate &amp;quot;came&amp;quot; .</Paragraph>
    <Paragraph position="1"> Predicate - the given predicate lemma Voice - whether the predicate is realized as an active or passive construction (Pradhan et al., 2004, claim approximately 11% of the sentences in PropBank use a passive instantiation) Phrase Type - the syntactic category (NP, PP, S, etc.) of the phrase corresponding to the semantic argument Distance - the relative displacement from the predicate, measured in intervening constituents (negative if the constituent appears prior to the predicate, positive if it appears after it) Head Word - the syntactic head of the phrase, calculated by finding the last noun of a Noun Phrase Path - the syntactic path through the parse tree, from the parse constituent to the predicate being classified (for example, in Figure 1, the path from Arg0 - &amp;quot;The officer&amp;quot; to the predicate &amp;quot;came&amp;quot;, is represented with the string NP |S |VP |VBD&amp;quot; represent upward and downward movements in the tree respectively) Preposition - the preposition of an argument in a PP, such as &amp;quot;during&amp;quot;, &amp;quot;at&amp;quot;, &amp;quot;with&amp;quot;, etc (for example, in Figure 1, the preposition for the PP with Argm-Loc label is &amp;quot;to&amp;quot;).</Paragraph>
    <Paragraph position="2"> In addition, an actor heuristic is adopted: where an instance can be labeled as A0 (actor) only if the argument is a subject before the predicate in active voice, or if the preposition &amp;quot;by&amp;quot; appears prior to this argument but after the predicate in a passive voice sentence. For example, if there is a set of labels, A0 (subject or actor) V (active) A0 (non actor), then the latter &amp;quot;A0&amp;quot; after V is skipped and labeled to another suitable role by this heuristic; such as the label with the second highest probability for this argument according to the PML estimate, or with the second shortest distance estimate by kNN.</Paragraph>
    <Section position="1" start_page="180" end_page="180" type="sub_section">
      <SectionTitle>
2.1 k Nearest Neighbour (kNN) Algorithm
</SectionTitle>
      <Paragraph position="0"> One instance-based learning algorithm is k-Nearest Neighbour (kNN), which is suitable when 1) instances can be mapped to points/classifications in n-dimensional feature dimension, 2) fewer than 20 features are utilized, and 3) training data is sufficiently abundant.</Paragraph>
      <Paragraph position="1"> One advantage of kNN is that training is very fast; one disadvantage is it is generally slow at testing. The implementation of kNN is described as following</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>