<?xml version="1.0" standalone="yes"?>
<Paper uid="J95-3002">
  <Title>Robust Learning, Smoothing, and Parameter Tying on Syntactic Ambiguity Resolution</Title>
  <Section position="10" start_page="345" end_page="347" type="concl">
    <SectionTitle>
7. Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> An integrated scoring function capable of incorporating various knowledge sources to resolve syntactic ambiguity problems is explored in this paper. In the baseline model, the parameters are estimated by using the maximum likelihood method. The MLE  Summary of performance for the Lex(L2)+Syn(L2) model using various performance enhancement methods. Values in parentheses correspond to performance excluding unambiguous sentences.</Paragraph>
    <Paragraph position="1">  approach fails to achieve satisfactory performance because the discrimination and robustness issues are not considered in the estimation process. To improve performance, a discrimination- and robustness-oriented method is adopted to directly pursue the correct ranking orders of possible alternative syntactic structures. In addition, this learning procedure is able to resolve problems resulting from statistical variations between the training corpus and real tasks.</Paragraph>
    <Paragraph position="2"> The effects of parameter smoothing for null events with Turing's formula and the Back-Off method are investigated in this paper. A better initial estimate of the parameters makes the robust learning procedure achieve better performance when many local optima exist in the parameter space. Significant improvement of 34.3% error reduction rate is attained when we apply the robust learning procedure on the smoothed parameters.</Paragraph>
    <Paragraph position="3"> Finally, a parameter tying scheme for rare events is proposed so that the unreliably estimated parameters are tied and trained together through the robust learning procedure. Thus, this approach makes it possible to tune all the parameters through the learning process. In addition, the number of parameters is significantly reduced with the tying process. The reduction of the number of parameters is over 99% for each language model. Moreover, the accuracy rate of 70.3% for parse tree selection, or 36.7% error reduction rate, is obtained by using this novel approach.</Paragraph>
    <Paragraph position="4"> To explore the areas for further improving the system, the remaining errors have been examined. It was found that a very large portion of errors result from attachment problems, including prepositional phrase (PP) attachment and modification scope for adverbial phrases, adjective phrases, and relative clauses, while less than 10% of the errors arise because of incorrect part-of-speech tagging. To further improve the lexical scoring module, some refinement mechanisms developed for our part-of-speech tagger (Lin, Chiang, and Su 1994) will be incorporated into this system. As for the attachment problems, we found that the system appears to have a preference for local attachment, which is not always inappropriate. The current model fails to deal with such problems because only syntactic information from two left contextual nonterminal symbols is consulted for computation. To resolve the attachment problems, integrating seman- null Computational Linguistics Volume 21, Number 3 tic information, such as word sense collocations, would be required. In addition, to enable the system to take into account information associated with long-distance dependency, we plan to modify the syntactic model so that it can evaluate structural dependency across various subtrees in the parse history. A large number of parameters will inevitably be required for such a formulation, and a large training corpus is thus needed for training. A bootstrapping procedure for parameter estimation with respect to a very large corpus, therefore, will be applied in future research.</Paragraph>
  </Section>
class="xml-element"></Paper>