<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1023">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Trace Prediction and Recovery With Unlexicalized PCFGs and Slash Features</Title>
  <Section position="8" start_page="183" end_page="183" type="concl">
    <SectionTitle>
6 Summary
</SectionTitle>
    <Paragraph position="0"> We presented an unlexicalized PCFG parser which applies a slash feature percolation mechanism to generate parse trees with empty elements and co-indexation of traces and fillers. The grammar was extracted from a version of the PENN tree-bank which was annotated with slash features and a set of other features that were added in order to improve the general parsing accuracy. The parser computes true Viterbi parses unlike most other parsers for treebank grammars which are not guaranteed to produce the most likely parse tree because they apply pruning strategies like beam search.</Paragraph>
    <Paragraph position="1"> We evaluated the parser using the standard PENN treebank training and test data. The labeled bracketing f-score of 86.6% is - to our knowledge - the best f-score reported for unlexicalized PCFGs, exceeding that of Klein and Manning (2003) by almost 1%. On the empty category prediction task, our parser outperforms the best previously reported system (Campbell, 2004) by 0.7% reaching an f-score of 84.1%, although the general parsing accuracy of our unlexicalized parser is 3% lower than that of the parser used by Campbell (2004). Our parser also ranks highest in terms of the co-indexation accuracy with 77.4% f-score, again outperforming the system of Campbell (2004) by 0.7%.</Paragraph>
  </Section>
class="xml-element"></Paper>