<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1513">
  <Title>Vancouver, October 2005. c(c)2005 Association for Computational Linguistics A Classifier-Based Parser with Linear Run-Time Complexity</Title>
  <Section position="3" start_page="0" end_page="125" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Two classifier-based deterministic dependency parsers for English have been proposed recently (Nivre and Scholz, 2004; Yamada and Matsumoto, 2003). Although they use different parsing algorithms, and differ on whether or not dependencies are labeled, they share the idea of greedily pursuing a single path, following parsing decisions made by a classifier. Despite their greedy nature, these parsers achieve high accuracy in determining dependencies. Although state-of-the-art statistical parsers (Collins, 1997; Charniak, 2000) are more accurate, the simplicity and efficiency of deterministic parsers make them attractive in a number of situations requiring fast, light-weight parsing, or parsing of large amounts of data. However, dependency analyses lack important information contained in constituent structures. For example, the tree-path feature has been shown to be valuable in semantic role labeling (Gildea and Palmer, 2002).</Paragraph>
    <Paragraph position="1"> We present a parser that shares much of the simplicity and efficiency of the deterministic dependency parsers, but produces both dependency and constituent structures simultaneously. Like the parser of Nivre and Scholz (2004), it uses the basic shift-reduce stack-based parsing algorithm, and runs in linear time. While it may seem that the larger search space of constituent trees (compared to the space of dependency trees) would make it unlikely that accurate parse trees could be built deterministically, we show that the precision and recall of constituents produced by our parser are close to those produced by statistical parsers with higher run-time complexity.</Paragraph>
    <Paragraph position="2"> One desirable characteristic of our parser is its simplicity. Compared to other successful approaches to corpus-based constituent parsing, ours is remarkably simple to understand and implement.</Paragraph>
    <Paragraph position="3"> An additional feature of our approach is its modularity with regard to the algorithm and the classifier that determines the parser's actions. This makes it very simple for different classifiers and different sets of features to be used with the same parser with very minimal work. Finally, its linear run-time complexity allows our parser to be considerably faster than lexicalized PCFG-based parsers. On the other hand, a major drawback of the classifier-based parsing framework is that, depending on  the classifier used, its training time can be much longer than that of other approaches.</Paragraph>
    <Paragraph position="4"> Like other deterministic parsers (and unlike many statistical parsers), our parser considers the problem of syntactic analysis separately from part-of-speech (POS) tagging. Because the parser greedily builds trees bottom-up in one pass, considering only one path at any point in the analysis, the task of assigning POS tags to words is done before other syntactic analysis. In this work we focus only on the processing that occurs once POS tagging is completed. In the sections that follow, we assume that the input to the parser is a sentence with corresponding POS tags for each word.</Paragraph>
  </Section>
class="xml-element"></Paper>