<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1005">
  <Title>Computing Consensus Translation from Multiple Machine Translation Systems Using Enhanced Hypotheses Alignment</Title>
  <Section position="6" start_page="37" end_page="38" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this work, we proposed a novel, theoretically well-founded procedure for computing a possibly new consensus translation from the outputs of multiple MT systems. In summary, the main con- null the BTEC Italian-English task through computing consensus translations from the output of two speech translation systems with different types of source language input.</Paragraph>
    <Paragraph position="1">  consensus a-d 28.5 25.0 58.9 tributions of this work compared to previous approaches are as follows: * The words of the original translation hypotheses are aligned in order to create a confusion network. The alignment procedure explicitly models word reordering.</Paragraph>
    <Paragraph position="2"> * A test corpus of translations generated by each of the systems is used for the unsupervised statistical alignment training. Thus, the decision on how to align two translations of a sentence takes the whole document context into account.</Paragraph>
    <Paragraph position="3"> * Large and significant gains in translation quality were obtained on various translation tasks and conditions.</Paragraph>
    <Paragraph position="4"> * A significant improvement of translation quality was achieved in a multi-source translation scenario. Here, we combined the output of MT systems which have different source and the same target language.</Paragraph>
    <Paragraph position="5"> * The proposed method can be effectively applied in speech translation in order to cope with the negative impact of speech recognition errors on translation accuracy.</Paragraph>
    <Paragraph position="6"> An important feature of a real-life application of the proposed alignment technique is that the lexicon and alignment probabilities can be updated with each translated sentence and/or text. Thus,  thecorrespondencebetweenwordsindifferenthypotheses and, consequently, the consensus translation can be improved overtime.</Paragraph>
  </Section>
class="xml-element"></Paper>