<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1057">
  <Title>Feedback Cleaning of Machine Translation Rules Using Automatic Evaluation</Title>
  <Section position="12" start_page="2" end_page="2" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we proposed a feedback cleaning method that utilizes automatic evaluation to remove incorrect/redundant translation rules. BLEU was utilized for the automatic evaluation of MT quality, and the hill-climbing algorithm was applied to searching for the combinatorial optimization. Utilizing features of this task, incorrect/redundant rules were removed from the initial solution, which contains all rules acquired from the training corpus. In addition, we proposed N-fold cross-cleaning to reduce the influence of the evaluation corpus size. Our experiments show that the MT quality was improved by 10% in paired comparison and by 0.045 in the BLEU score. This is considerable improvement over the previous methods.</Paragraph>
  </Section>
class="xml-element"></Paper>