<?xml version="1.0" standalone="yes"?>
<Paper uid="J00-1004">
  <Title>Learning Dependency Translation Models as Collections of Finite-State Head Transducers Hiyan Alshawi*</Title>
  <Section position="7" start_page="57" end_page="58" type="concl">
    <SectionTitle>
6. Concluding Remarks
</SectionTitle>
    <Paragraph position="0"> Formalisms for finite-state and context-free transduction have a long history (e.g., Lewis and Stearns 1968; Aho and Ullman 1972), and such formalisms have been applied to the machine translation problem, both in the finite-state case (e.g., Vilar et al.</Paragraph>
    <Paragraph position="1"> 1996) and the context-free case (e.g., Wu 1997). In this paper we have added to this line of research by providing a method for automatically constructing fully lexicalized statistical dependency transduction models from training examples.</Paragraph>
    <Paragraph position="2"> Automatically training a translation system brings important benefits in terms of maintainability, robustness, and reducing expert coding effort as compared with tra- null Alshawi, Bangalore, and Douglas Learning Dependency Translation Models ditional rule-based translation systems (a number of which are described in Hutchins and Somers \[1992\]). The reduction of effort results, in large part, from being able to do without artificial intermediate representations of meaning; we do not require the development of semantic mapping rules (or indeed any rules) or the creation of a corpus including semantic annotations. Compared with left-to-right transduction, middle-out transduction also aids robustness because, when complete derivations are not available, partial derivations tend to have meaningful headwords.</Paragraph>
    <Paragraph position="3"> At the same time, we believe our method has advantages over the approach developed initially at IBM (Brown et al. 1990; Brown et al. 1993) for training translation systems automatically. One advantage is that our method attempts to model the natural decomposition of sentences into phrases. Another is that the compilation of this decomposition into lexically anchored finite-state head transducers produces implementations that are much more efficient than those for the IBM model. In particular, our search algorithm finds optimal transductions of test sentences in less than &amp;quot;real time&amp;quot; on a 300MHz processor, that is, the time to translate an utterance is less than the time taken to speak it, an important consideration for our speech translation application. null</Paragraph>
  </Section>
class="xml-element"></Paper>