<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-0706">
  <Title>Exploring Evidence for Shallow Parsinga0</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Shallow parsing is studied as an alternative to full-sentence parsing. Rather than producing a complete analysis of sentences, the alternative is to perform only partial analysis of the syntactic structures in a text (Harris, 1957; Abney, 1991; Greffenstette, 1993). A lot of recent work on shallow parsing has been influenced by Abney's work (Abney, 1991), who has suggested to &amp;quot;chunk&amp;quot; sentences to base level phrases. For example, the sentence &amp;quot;He reckons the current account deficit will narrow to only $ 1.8 billion in September .&amp;quot; would be chunked as follows (Tjong Kim Sang and Buchholz, 2000): [NP He ] [VP reckons ] [NP the current account deficit ] [VP will narrow ] [PP a1This research is supported by NSF grants IIS-9801638, ITR-IIS-0085836 and an ONR MURI Award.</Paragraph>
    <Paragraph position="1"> to ] [NP only $ 1.8 billion ] [PP in ] [NP September] .</Paragraph>
    <Paragraph position="2"> While earlier work in this direction concentrated on manual construction of rules, most of the recent work has been motivated by the observation that shallow syntactic information can be extracted using local information -- by examining the pattern itself, its nearby context and the local part-of-speech information. Thus, over the past few years, along with advances in the use of learning and statistical methods for acquisition of full parsers (Collins, 1997; Charniak, 1997a; Charniak, 1997b; Ratnaparkhi, 1997), significant progress has been made on the use of statistical learning methods to recognize shallow parsing patterns -- syntactic phrases or words that participate in a syntactic relationship (Church, 1988; Ramshaw and Marcus, 1995; Argamon et al., 1998; Cardie and Pierce, 1998; Munoz et al., 1999; Punyakanok and Roth, 2001; Buchholz et al., 1999; Tjong Kim Sang and Buchholz, 2000).</Paragraph>
    <Paragraph position="3"> Research on shallow parsing was inspired by psycholinguistics arguments (Gee and Grosjean, 1983) that suggest that in many scenarios (e.g., conversational) full parsing is not a realistic strategy for sentence processing and analysis, and was further motivated by several arguments from a natural language engineering viewpoint.</Paragraph>
    <Paragraph position="4"> First, it has been noted that in many natural language applications it is sufficient to use shallow parsing information; information such as noun phrases (NPs) and other syntactic sequences have been found useful in many large-scale language processing applications including information extraction and text summarization (Grishman, 1995; Appelt et al., 1993). Second, while training a full parser requires a collection of fully parsed sentences as training corpus, it is possible to train a shallow parser incrementally. If all that is available is a collection of sentences annotated for NPs, it can be used to produce this level of analysis. This can be augmented later if more information is available. Finally, the hope behind this research direction was that this incremental and modular processing might result in more robust parsing decisions, especially in cases of spoken language or other cases in which the quality of the natural language inputs is low -- sentences which may have repeated words, missing words, or any other lexical and syntactic mistakes.</Paragraph>
    <Paragraph position="5"> Overall, the driving force behind the work on learning shallow parsers was the desire to get better performance and higher reliability. However, since work in this direction has started, a significant progress has also been made in the research on statistical learning of full parsers, both in terms of accuracy and processing time (Charniak, 1997b; Charniak, 1997a; Collins, 1997; Ratnaparkhi, 1997).</Paragraph>
    <Paragraph position="6"> This paper investigates the question of whether work on shallow parsing is worthwhile. That is, we attempt to evaluate quantitatively the intuitions described above -- that learning to perform shallow parsing could be more accurate and more robust than learning to generate full parses. We do that by concentrating on the task of identifying the phrase structure of sentences -- a byproduct of full parsers that can also be produced by shallow parsers. We investigate two instantiations of this task, &amp;quot;chucking&amp;quot; and identifying atomic phrases. And, to study robustness, we run our experiments both on standard Penn Treebank data (part of which is used for training the parsers) and on lower quality data -- the Switchboard data.</Paragraph>
    <Paragraph position="7"> Our conclusions are quite clear. Indeed, shallow parsers that are specifically trained to perform the tasks of identifying the phrase structure of a sentence are more accurate and more robust than full parsers. We believe that this finding, not only justifies work in this direction, but may even suggest that it would be worthwhile to use this methodology incrementally, to learn a more complete parser, if needed.</Paragraph>
  </Section>
class="xml-element"></Paper>