<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2089">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics A Best-First Probabilistic Shift-Reduce Parser</Title>
  <Section position="5" start_page="693" end_page="693" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> As mentioned in section 2, our parsing approach can be seen as an extension of the approach of Sagae and Lavie (2005). Sagae and Lavie evaluated their deterministic classifier-based parsing framework using two classifiers: support vector machines (SVM) and k-nearest neighbors (kNN).</Paragraph>
    <Paragraph position="1"> Although the kNN-based parser performed poorly, the SVM-based parser achieved about 86% precision and recall (or 87.5% using gold-standard POS tags) on the WSJ test section of the Penn Treebank, taking only 11 minutes to parse the test set. Sagae and Lavie's parsing algorithm is similar to the one used by Nivre and Scholz (2004) for deterministic dependency parsing (using kNN). Yamada and Matsumoto (2003) have also presented a deterministic classifier-based (SVM-based) dependency parser, but using a different parsing algorithm, and using only unlabeled dependencies.</Paragraph>
    <Paragraph position="2"> Tsuruoka and Tsujii (2005) developed a classifier-based parser that uses the chunk-parsing algorithm and achieves extremely high parsing speed, but somewhat low recall. The algorithm  is based on reframing the parsing task as several sequential chunking tasks.</Paragraph>
    <Paragraph position="3"> Finally, our parser is in many ways similar to the parser of Ratnaparkhi (1997). Ratnaparkhi's parser uses maximum-entropy models to determine the actions of a parser based to some extent on the shift-reduce framework, and it is also capable of pursuing several paths and returning the top-n highest scoring parses for a sentence. However, in addition to using different features for parsing, Ratnaparkhi's parser uses a different, more complex algorithm. The use of a more involved algorithm allows Ratnaparkhi's parser to work with arbitrary branching trees without the need of the binarization transform employed here. It breaks the usual reduce actions into smaller pieces (CHECK and BUILD), and uses two separate passes (not including the part-of-speech tagging pass) for determining chunks and higher syntactic structures separately. Instead of keeping a stack, the parser makes multiple passes over the input string, like the dependency parsing algorithm used by Yamada and Matsumoto. Our parser, on the other hand, uses a simpler stack-based shift-reduce (LRlike) algorithm for trees with only unary and binary productions.</Paragraph>
  </Section>
class="xml-element"></Paper>