<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2902">
  <Title>Porting Statistical Parsers with Data-Defined Kernels</Title>
  <Section position="8" start_page="11" end_page="12" type="relat">
    <SectionTitle>
6 Related Work
</SectionTitle>
    <Paragraph position="0"> Most research in the field of parsing has focused on the Wall Street Journal corpus. Several researchers have addressed the portability of these WSJ parsers to other domains, but mostly without addressing the issue of how a parser can be designed specifically for porting to another domain. Unfortunately, no direct empirical comparison is possible between our results and results with other parsers, because there is no standard portability benchmark to date where a small amount of data from a target domain is used.</Paragraph>
    <Paragraph position="1"> (Ratnaparkhi, 1999) performed portability experiments with a Maximum Entropy parser and demonstrated that the parser trained on WSJ achieves far worse results on the Brown corpus sections. Adding a small amount of data from the target domain improves the results, but accuracy is still much lower than the results on the WSJ. They reported results when their parser was trained on the WSJ training  set plus a portion of 2,000 sentences from a Brown corpus section. They achieved 80.9%/80.3% recall/precision for section K, and 80.6%/81.3% for section N.7 Our analogous method (TOP-Focus) achieved much better accuracy (3.7% and 4.9% better F1, respectively).</Paragraph>
    <Paragraph position="2"> In addition to portability experiments with the parsing model of (Collins, 1997), (Gildea, 2001) provided a comprehensive analysis of parser portability. On the basis of this analysis, a technique for parameter pruning was proposed leading to a significant reduction in the model size without a large decrease of accuracy. Gildea (2001) only reports results on sentences of 40 or less words on all the Brown corpus sections combined, for which he reports 80.3%/81.0% recall/precision when training only on data from the WSJ corpus, and 83.9%/84.8% when training on data from the WSJ corpus and all sections of the Brown corpus.</Paragraph>
    <Paragraph position="3"> (Roark and Bacchiani, 2003) performed experiments on supervised and unsupervised PCFG adaptation to the target domain. They propose to use the statistics from a source domain to define priors over weights. However, in their experiments they used only trivial sub-cases of this approach, namely, count merging and model interpolation.</Paragraph>
    <Paragraph position="4"> They achieved very good improvement over their baseline and over (Gildea, 2001), but the absolute accuracies were still relatively low (as discussed above). They report results with combined Brown data (on sentences of 100 words or less), achieving 81.3%/80.9% when training only on the WSJ corpus and 85.4%/85.9% with their best method using the data from both domains.</Paragraph>
  </Section>
class="xml-element"></Paper>