<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1516">
  <Title>Strictly Lexical Dependency Parsing</Title>
  <Section position="3" start_page="0" end_page="152" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> There has been a great deal of progress in statistical parsing in the past decade (Collins, 1996; Collins, 1997; Chaniak, 2000). A common characteristic of these parsers is their use of lexicalized statistics. However, it was discovered recently that bi-lexical statistics (parameters that involve two words) actually played much smaller role than previously believed. It was found in (Gildea, 2001) that the removal of bi-lexical statistics from a state-of-the-art PCFG parser resulted very small change in the output. Bikel (2004) observed that the bi-lexical statistics accounted for only 1.49% of the bigram statistics used by the parser. When considering only bigram statistics involved in the highest probability parse, this percentage becomes 28.8%. However, even when the bi-lexical statistics do get used, they are remarkably similar to their back-off values using part-of-speech tags.</Paragraph>
    <Paragraph position="1"> Therefore, the utility of bi-lexical statistics becomes rather questionable. Klein and Manning (2003) presented an unlexicalized parser that eliminated all lexicalized parameters. Its performance was close to the state-of-the-art lexicalized parsers.</Paragraph>
    <Paragraph position="2"> We present a statistical dependency parser that represents the other end of spectrum where all statistical parameters are lexical and the parser does not require part-of-speech tags or grammatical categories. We call this strictly lexicalized parsing.</Paragraph>
    <Paragraph position="3"> A part-of-speech lexicon has always been considered to be a necessary component in any natural language parser. This is true in early rule-based as well as modern statistical parsers and in dependency parsers as well as constituency parsers.</Paragraph>
    <Paragraph position="4"> The need for part-of-speech tags arises from the sparseness of natural language data. They provide generalizations of words that are critical for parsers to deal with the sparseness. Words belonging to the same part-of-speech are expected to have the same syntactic behavior.</Paragraph>
    <Paragraph position="5"> Instead of part-of-speech tags, we rely on distributional word similarities computed automatically from a large unannotated text corpus. One of the benefits of strictly lexicalized parsing is that  fundsinvestors continue to pour cash into moneyMany[?] 01 2 3 456 78 9 the parser can be trained with a treebank that only contains the dependency relationships between words. The annotators do not need to annotate parts-of-speech or non-terminal symbols (they don't even have to know about them), making the construction of the treebank easier.</Paragraph>
    <Paragraph position="6"> Strictly lexicalized parsing is especially beneficial for languages such as Chinese, where parts-of-speech are not as clearly defined as English. In Chinese, clear indicators of a word's part-of-speech such as suffixes -ment, -ous or function words such as the, are largely absent. In fact, monolingual Chinese dictionaries that are mainly intended for native speakers almost never contain part-of-speech information.</Paragraph>
    <Paragraph position="7"> In the next section, we present a method for modeling the probabilities of dependency trees.</Paragraph>
    <Paragraph position="8"> Section 3 applies similarity-based smoothing to the probability model to deal with data sparseness.</Paragraph>
    <Paragraph position="9"> We then present experimental results with the Chinese Treebank in Section 4 and discuss related work in Section 5.</Paragraph>
  </Section>
class="xml-element"></Paper>