<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1627">
  <Title>Efficient Search for Inversion Transduction Grammar</Title>
  <Section position="8" start_page="230" end_page="230" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> A* search for Viterbi alignment selection under ITG is ef cient using IBM Model 1 as an outside estimate. The experimental results indicate that despite being a more relaxed word-for-word alignment model than ITG, IBM Model 1 can serve as an ef cient and reliable approximation of ITG in terms of Viterbi alignment probability. This is more true when we apply Model 1 to both translation directions and take the minimum of both.</Paragraph>
    <Paragraph position="1"> We have also tried to incorporate estimates of binary rule probabilities to make the outside estimate even sharper. However, the further improvement was marginal.</Paragraph>
    <Paragraph position="2"> We are able to nd the ITG Viterbi translation using our A* decoding algorithm with an outside estimate that combines outside bigrams and translation probabilities for outside words. The hook trick gave us a signi cant further speedup; we believe this to be the rst demonstrated practical application of this technique.</Paragraph>
    <Paragraph position="3"> Interestingly, the BLEU score for the optimal translations under the probabilistic model is lower than we achieve with our best bigram-based system using pruning. However, this system makes use of the A* heuristic, and our speed/performance curve shows that the heuristic allows us to achieve higher BLEU scores with the same amount of computation. In the case of trigram integrated decoding, there is 1 point of BLEU score improvement by moving from a typical CYK plus beam search decoder to a decoder using A* plus beam search.</Paragraph>
    <Paragraph position="4"> However, without knowing what words will appear in the output language, a very sharp outside estimate to further bring down the number of combinations is dif cult to achieve.</Paragraph>
    <Paragraph position="5"> The brighter side of the move towards optimal decoding is that the A* search strategy leads us to the region of the search space that is close to the optimal result, where we can more easily nd good translations.</Paragraph>
    <Paragraph position="6"> Acknowledgments This work was supported by NSF ITR IIS-09325646 and NSF ITR IIS0428020. null</Paragraph>
  </Section>
class="xml-element"></Paper>