<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-3004">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Annotation Schemes and their Influence on Parsing Results</Title>
  <Section position="7" start_page="23" end_page="23" type="concl">
    <SectionTitle>
5 Conclusions and Outlook
</SectionTitle>
    <Paragraph position="0"> We presented an analysis of the influences of the particularities of annotation schemes on parsing results via a comparison of two German treebanks, NeGra and T&amp;quot;uBa-D/Z, based on a step-wise approximation of both treebanks. The experiments show that as treebanks are approximated, the parsing results also get closer. When annotation structure is deleted in T&amp;quot;uBa-D/Z, the number of crossing brackets drops, but F-Measure drops, too. When annotation structure is added in Ne-Gra, the contrary happens. We can conclude that, being interested in good F-Measure results, the deep T&amp;quot;uBa-D/Z structures are more appropriate for parsing than NeGra's flat structures. Moreover, we have observed that it is beneficial to provide the parser with the gold POS tags at parsing time.</Paragraph>
    <Paragraph position="1"> However, we see that especially when parsing with grammatical functions, data sparseness becomes a serious problem, making the results less reliable.</Paragraph>
    <Paragraph position="2"> Seen in the context of a parse tree, the expansion probability of a PCFG rule just covers a subtree of height 1. This is a clear deficiency of PCFGs since this way, e.g., the expansion probability of a VP is independent of the choice of the verb. Our future work will start at this point. We will conduct further experiments with the Stanford Parser (Klein and Manning, 2003) which considers broader contexts in its probability. It uses markovization to reduce horizontal context (right hand sides of rules are broken up) and add vertical context (rule probabilities are conditioned on (grand-)parent-node information). This way, we expect further insights in NeGra's an T&amp;quot;uBa-D/Z's annotation schemes.</Paragraph>
  </Section>
class="xml-element"></Paper>