<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-2026">
  <Title>ing of a tree adjoining grammar using finitestate machines. In Proceedings of the Sixth International Workshop on tree Adjoining</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Experimental Results
</SectionTitle>
    <Paragraph position="0"> We present results in which our classes are defined entirely with respect to syntactic behavior. This is because we do not have available an important corpus annotated with semantics.</Paragraph>
    <Paragraph position="1"> We train on the Wall Street Journal (WSJ) corpus. We evaluate by taking a list of 205 sentences which are chosen at random from entries to WORDSEYE made by the developers (who were testing the graphical component using a different parser). Their average length is 6.3 words. We annotated the sentences by hand for the desired dependency structure, and then compared the structural output of PARSLI to the gold standard (we disregarded the functional and semantic annotations produced by PARSLI). We evaluate performance using accuracy, the ration of  the number of dependency arcs which are correctly found (same head and daughter nodes) in the best parse for each sentence to the number of arcs in the entire test corpus. We also report the percentage of sentences for which we find the correct dependency tree (correctness). For our test corpus, we obtain an accuracy of 0.78 and a correctness of 0.58. The average transduction time per sentence (including initialization of the parser) is 0.29 s. Figure 5 shows the dependence of the scores on sentence length. As expected, the longer the sentence, the worse the score.</Paragraph>
    <Paragraph position="2"> We can obtain the n-best paths through the FST; the scores for n-best paths are summarized in Figure 6. Since the scores keep increasing, we believe that we can further improve our 1-best results by better choosing the correct path. We intend to adapt the FSTs to use probabilities of attaching particular supertags to other supertags (rather than uniform weights for all attachments) in order to better model the probability of different analyses. Another option, of course, is bilexical probabilities.</Paragraph>
  </Section>
class="xml-element"></Paper>