<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1100">
  <Title>Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 795-802, Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Morphology and Reranking for the Statistical Parsing of Spanish</Title>
  <Section position="2" start_page="0" end_page="795" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Initial methods for statistical parsing were mainly developed through experimentation on English data sets. Subsequent research has focused on applying these methods to other languages. There has been widespread evidence that new languages exhibit linguistic phenomena that pose considerable challenges to techniques originally developed for English; because of this, an important area of current research concerns how to model these phenomena more accurately within statistical approaches. In this paper, we investigate this question within the context of parsing Spanish. We describe two methods for incorporating detailed features in a Spanish parser, building on a baseline model that is a lexicalized PCFG originally developed for English.</Paragraph>
    <Paragraph position="1"> Our first model uses morphology to improve the performance of the baseline model. English is a morphologically-impoverished language, while most of the world's languages exhibit far richer morphologies. Spanish is one of these languages. For instance, the forms of Spanish nouns, determiners, and adjectives reflect both number and gender; pronouns reflect gender, number, person, and case. Furthermore, morphological constraints may be manifested at the syntactic level: certain constituents of a noun phrase are constrained to agree in number and gender, and a verb is constrained to agree in number and person with its subject. Hence, morphology gives us important structural cues about how the words in a Spanish sentence relate to one another.</Paragraph>
    <Paragraph position="2"> The mechanism we employ for incorporating morphology into the PCFG model (the Model 1 parser in (Collins, 1999)) is the modification of its part-of-speech (POS) tagset; in this paper, we explain how this mechanism allows the parser to better capture morphological constraints.</Paragraph>
    <Paragraph position="3"> All of the experiments in this paper are carried out using a freely-available Spanish treebank produced by the 3LB project (Navarro et al., 2003).</Paragraph>
    <Paragraph position="4"> This resource contains around 3,500 hand-annotated trees encoding ample morphological information.</Paragraph>
    <Paragraph position="5"> We could not use all of this information and adequately train the resulting parameters due to limited training data. Hence, we used development data to test the performance of several models, each incorporating a subset of morphological information. The highest-accuracy model on the development set uses the mode and number of verbs, as well as the number of adjectives, determiners, nouns, and pronouns. On test data, it reaches F1 accuracy of 83.6%/83.9%/79.4% for labeled constituents, unlabeled dependencies, and labeled dependencies, respectively. The baseline model, which makes almost no use of morphology, achieves 81.2%/82.5%/77.0% in these same measures.</Paragraph>
    <Paragraph position="6"> We use the morphological model from the aforementioned experiments as a base parser in a second set of experiments. Here we investigate the efficacy of a reranking approach for parsing Spanish by using  arbitrary structural features. Previous work in statistical parsing (Collins and Koo, 2005) has shown that applying reranking techniques to the n-best output of a base parser can improve parsing performance. Applying an exponentiated gradient reranking algorithm (Bartlett et al., 2004) to the n-best output of our morphologically-informed Spanish parsing model gives us similar improvements. Using the reranking model combined with the morphological model raises performance to 85.1%/84.7%/80.2% F1 accuracy for labeled constituents, unlabeled dependencies, and labeled dependencies.</Paragraph>
  </Section>
class="xml-element"></Paper>