<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1408">
  <Title>An Efficient A* Search Algorithm for Statistical Machine Translation</Title>
  <Section position="9" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have developed sophisticated admissible and almost admissible heuristic functions for statistical machine translation. We have focussed on Model 4, but most of the computations could be easily extended to other statistical alignment models (like HMM or Model 5). We especially have observed the following effects: The heuristic function has a strong effect on the efficiency of the A* search. Without any heuristic function only 75 % of the test corpus sentences can be translated (using the  best admissible heuristic function TFLD we can translate 82 %.</Paragraph>
    <Paragraph position="1"> Using the empirical heuristic function we can translate 96 % of the sentences with A* search. This heuristic function does not guarantee to avoid search errors, but this case never occurred in our experiments.</Paragraph>
    <Paragraph position="2"> From these results we conclude that it is often possible to faster compute acceptable results using a beam search approach. Therefore, this is the method of choice in practice. From a theoretical viewpoint it is interesting that using A* it is possible to translate guaranteed without search errors. In addition, without having a chance to perform search without search errors it is almost impossible to assess if errors in translation should be assigned to the model/training or to the search heuristics. Therefore, the A* algorithm is especially useful during the development of a statistical machine translation system.</Paragraph>
  </Section>
class="xml-element"></Paper>