<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1036">
  <Title>Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 281-290, Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Compiling Comp Ling: Practical Weighted Dynamic Programming and the Dyna Language[?]</Title>
  <Section position="2" start_page="0" end_page="281" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In this paper, we generalize some modern probabilistic parsing techniques to a broader class of weighted deductive algorithms. Our implemented system encapsulates these implementation techniques behind a clean interface--a small high-level specification language, Dyna, which compiles into C++ classes. This system should help the HLT community to experiment more easily with new models and algorithms.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
1.1 Dynamic programming as deduction
</SectionTitle>
      <Paragraph position="0"> The &amp;quot;parsing as deduction&amp;quot; framework (Pereira and Warren, 1983) is now over 20 years old. It provides an elegant notation for specifying a variety of parsing algorithms (Shieber et al., 1995), including algorithms for probabilistic or other semiring-weighted parsing (Goodman, 1999). In the parsing community, new algorithms are often stated simply as a set of deductive inference rules (Sikkel, 1997; Eisner and Satta, 1999).</Paragraph>
      <Paragraph position="1"> It is also straightforward to specify other NLP algorithms this way. Syntactic MT models, language models, and stack decoders can be easily described using deductive rules. So can operations on finite-state and infinite-state machines.</Paragraph>
      <Paragraph position="2"> [?]We thank Joshua Goodman, David McAllester, and Paul Ruhlen for useful early discussions; pioneer users Markus Dreyer, David Smith, and Roy Tromble for their feedback and input; John Blatz for discussion of program transformations; and several reviewers for useful criticism. This work was supported by NSF ITR grant IIS-0313193, ONR MURI grant N00014-01-1-0685, and a Hertz Foundation fellowship to the third author. The views expressed are not necessarily endorsed by the sponsors.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="281" type="sub_section">
      <SectionTitle>
1.2 The role of toolkits
</SectionTitle>
      <Paragraph position="0"> One might regard deductive inference as merely a helpful perspective for teaching old algorithms and thinking about new ones, linking NLP to logic and classical AI. Real implementations would then be carefully hand-coded in a traditional language.</Paragraph>
      <Paragraph position="1"> That was the view ten years ago of finite-state machines--that FSMs were part of the theoretical backbone of CL, linking the field to the theory of computation. Starting in the mid-1990's, however, finite-state methods came to the center of applied NLP as researchers at Xerox, AT&amp;T, Groningen and elsewhere improved the expressive power of FSMs by moving from automata to transducers, adding semiring weights, and developing powerful new regular-expression operators and algorithms for these cases. They also developed software. Karttunen et al. (1996) built an FSM toolkit that allowed construction of morphological analyzers for many languages. Mohri et al. (1998) built a weighted toolkit that implemented novel algorithms (e.g., weighted minimization, on-the-fly composition) and scaled up to handle large-vocabulary continuous ASR. At the same time, renewed community-wide interest in shallow methods for information extraction, chunking, MT, and dialogue processing meant that such off-the-shelf FS toolkits became the core of diverse systems used in cutting-edge research.</Paragraph>
      <Paragraph position="2"> The weakness of FSMs, of course, is that they are only finite-state. One would like something like AT&amp;T's FSM toolkit that also handles the various formalisms now under consideration for lexicalized grammars, non-context-free grammars, and syntax-based MT--and hold the promise of extending to other formalisms and applications not yet imagined.</Paragraph>
      <Paragraph position="3"> We believe that deductive inference should play the role of regular expressions and FSMs, providing the theoretical foundation for such an effort. Many engineering ideas in the field can be regarded, we  1. :- double item=0. % declares that all item values are doubles, default is 0 2. constit(X,I,K) += rewrite(X,W) * word(W,I,K). % a constituent is either a word . . . 3. constit(X,I,K) += rewrite(X,Y,Z) * constit(Y,I,J) * constit(Z,J,K). % . . . or a combination of two adjacent subconstituents 4. goal += constit(&amp;quot;s&amp;quot;,0,N) whenever ?ends at(N). % a parse is any s constituent that covers the input string  believe, as ideas for how to specify, transform, or compile systems of inference rules.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>