<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1515">
  <Title>Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Constituent Parsing by Classification</Title>
  <Section position="3" start_page="0" end_page="141" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Discriminative machine learning methods have improved accuracy on many NLP tasks, such as POS-tagging (Toutanova et al., 2003), machine translation (Och &amp; Ney, 2002), and relation extraction (Zhao &amp; Grishman, 2005). There are strong reasons to believe the same would be true of parsing. However, only limited advances have been made thus far, perhaps due to various limitations of extant discriminative parsers. In this paper, we present some innovations aimed at reducing or eliminating some of these limitations, specifically for the task of constituent parsing: null * We show how constituent parsing can be performed using standard classification techniques.</Paragraph>
    <Paragraph position="1"> * Classifiers for different non-terminal labels can be induced independently and hence training can be parallelized.</Paragraph>
    <Paragraph position="2"> * The parser can use arbitrary information to evaluate candidate constituency inferences.</Paragraph>
    <Paragraph position="3"> * Arbitrary confidence scores can be aggregated in a principled manner, which allows beam search.</Paragraph>
    <Paragraph position="4"> In Section 2 we describe our approach to parsing. In Section 3 we present experimental results.</Paragraph>
    <Paragraph position="5"> The following terms will help to explain our work.</Paragraph>
    <Paragraph position="6"> A span is a range over contiguous words in the input sentence. Spans cross if they overlap but neither contains the other. An item (or constituent) is a (span,label) pair. A state is a set of parse items, none of which may cross. A parse inference is a pair (S,i), given by the current state S and an item i to be added to it. A parse path (or history) is a sequence of parse inferences over some input sentence (Klein &amp; Manning, 2001). An item ordering (ordering, for short) constrains the order in which items may be inferred. In particular, if we prescribe a complete item ordering, the parser is deterministic (Marcus, 1980) and each state corresponds to a unique parse path.</Paragraph>
    <Paragraph position="7"> For some input sentence and gold-standard parse, a state is correct if the parser can infer zero or more additional items to obtain the gold-standard parse. A parse path is correct if it leads to a correct state. An  inference is correct if adding its item to its state is correct.</Paragraph>
  </Section>
class="xml-element"></Paper>