<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1021">
  <Title>Multilingual Dependency Parsing using Bayes Point Machines</Title>
  <Section position="3" start_page="0" end_page="160" type="intro">
    <SectionTitle>
2 Data
</SectionTitle>
    <Paragraph position="0"> We utilize publicly available resources in Arabic, Chinese, Czech, and English for training our dependency parsers.</Paragraph>
    <Paragraph position="1"> For Czech we used the Prague Dependency Tree-bank version 1.0 (LDC2001T10). This is a corpus of approximately 1.6 million words. We divided the data into the standard splits for training, devel- null opment test and blind test. The Prague Czech Dependency Treebank is provided with human-edited and automatically-assigned morphological information, including part-of-speech labels. Training and evaluation was performed using the automatically-assigned labels.</Paragraph>
    <Paragraph position="2"> For Arabic we used the Prague Arabic Dependency Treebank version 1.0 (LDC2004T23).</Paragraph>
    <Paragraph position="3"> Since there is no standard split of the data into training and test sections, we made an approximate 70%/15%/15% split for training/development test/blind test by sampling whole files. The Arabic Dependency Treebank is considerably smaller than that used for the other languages, with approximately 117,000 tokens annotated for morphological and syntactic relations. The relatively small size of this corpus, combined with the morphological complexity of Arabic and the heterogeneity of the corpus (it is drawn from five different newspapers across a three-year time period) is reflected in the relatively low dependency accuracy reported below.</Paragraph>
    <Paragraph position="4"> As with the Czech data, we trained and evaluated using the automatically-assigned part-of-speech labels provided with the data.</Paragraph>
    <Paragraph position="5"> Both the Czech and the Arabic corpora are annotated in terms of syntactic dependencies. For English and Chinese, however, no corpus is available that is annotated in terms of dependencies. We therefore applied head-finding rules to treebanks that were annotated in terms of constituency.</Paragraph>
    <Paragraph position="6"> For English, we used the Penn Treebank version 3.0 (Marcus et al., 1993) and extracted dependency relations by applying the head-finding rules of (Yamada and Matsumoto, 2003). These rules are a simplification of the head-finding rules of (Collins, 1999). We trained on sections 02-21, used section 24 for development test and evaluated on section 23. The English Penn Treebank contains approximately one million tokens. Training and evaluation against the development test set was performed using human-annotated part-of-speech labels. Evaluation against the blind test set was performed using part-of-speech labels assigned by the tagger described in (Toutanova et al., 2003).</Paragraph>
    <Paragraph position="7"> For Chinese, we used the Chinese Treebank version 5.0 (Xue et al., 2005). This corpus contains approximately 500,000 tokens. We made an approximate 70%/15%/15% split for training/development test/blind test by sampling whole files. As with the English Treebank, training and evaluation against the development test set was performed using human-annotated part-of-speech labels. For evaluation against the blind test section, we used an implementation of the tagger described in (Toutanova et al., 2003). Trained on the same training section as that used for training the parser and evaluated on the development test set, this tagger achieved a token accuracy of 92.2% and a sentence accuracy of 63.8%.</Paragraph>
    <Paragraph position="8"> The corpora used vary in homogeneity from the extreme case of the English Penn Treebank (a large corpus drawn from a single source, the Wall Street Journal) to the case of Arabic (a relatively small corpus-approximately 2,000 sentences-drawn from multiple sources). Furthermore, each language presents unique problems for computational analysis. Direct comparison of the dependency parsing results for one language to the results for another language is therefore difficult, although we do attempt in the discussion below to provide some basis for a more direct comparison. A common question when considering the deployment of a new language for machine translation is whether the natural language components available are of sufficient quality to warrant the effort to integrate them into the machine translation system. It is not feasible in every instance to do the integration work first and then to evaluate the output.</Paragraph>
    <Paragraph position="9"> Table 1 summarizes the data used to train the parsers, giving the number of tokens (excluding traces and other empty elements) and counts of sentences. null</Paragraph>
  </Section>
class="xml-element"></Paper>