<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1041">
  <Title>Using Probabilistic Models as Predictors for a Symbolic Parser</Title>
  <Section position="9" start_page="326" end_page="327" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented an architecture for the fusion of information contributed from a variety of components which are either based on expert knowledge or have been trained on quite different data collections. The results of the experiments show that there is a high degree of synergy between these different contributions, even if they themselves are fairly unreliable. Integrating all the available predictors we were able to improve the overall labelled accuracy on a standard test set for German to 91.1%, a level which is as least as good as the results reported for alternative approaches to parsing German.</Paragraph>
    <Paragraph position="1"> The result we obtained also challenges the common perception that rule-based parsers are necessarily inferior to stochastic ones. Supplied with appropriate helper components, the WCDG parser not only reached a surprisingly high level of output quality but in addition appears to be fairly stable against changes in the text type it is applied to (Foth et al., 2005).</Paragraph>
    <Paragraph position="2"> We attribute the successful integration of different information sources primarily to the fundamental ability of the WCDG grammar to combine evidence in a soft manner. If unreliable information needs to be integrated, this possibility is certainly an undispensible prerequisite for preventing local errors from accumulating and leading to an unacceptably low degree of reliability for the whole system eventually. By integrating the different predictors into the WCDG parsers's general mechanism for evidence arbitration, we not only avoided the adverse effect of individual error rates multiplying out, but instead were able to even raise the degree of output quality substantially.</Paragraph>
    <Paragraph position="3"> From the fact that the combination of all predictor components achieved the best results, even if the individual predictions are fairly unreliable, we can also conclude that diversity in the selection of predictor components is more important than the reliability of their contributions. Among the available predictor components which could be integrated into the parser additionally, the approach of (McDonald et al., 2005) certainly looks most promising. Compared to the shift-reduce parser which has been used as one of the predictor components for our experiments, it seems particularly attractive because it is able to predict non-projective structures without any additional provision, thus avoiding the misfit between our (non-projective) gold standard annotations and the restriction to projective structures that our shift-reduce parser suffers from.</Paragraph>
    <Paragraph position="4"> Another interesting goal of future work might be to even consider dynamic predictors, which can change their behaviour according to text type and perhaps even to text structure. This, however, would also require extending and adapting the cur- null rently dominating standard scenario of parser evaluation substantially.</Paragraph>
  </Section>
class="xml-element"></Paper>