<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-1010">
  <Title>Greedy Decoding for Statistical Machine Translation in Almost Linear Time</Title>
  <Section position="6" start_page="550" end_page="550" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we have analyzed the complexity of the greedy decoding algorithm originally presented in Germann et al. (2001) and presented improvements that drastically reduce the decoder's complexity and speed to practically linear time.</Paragraph>
    <Paragraph position="1">  a18 decoding (with 10 translations per input word considered, a list of 498 candidates for INSERT, a maximum swap distance of 2 and a maximum swap segment size of 5). The profiles shown are cumulative, so that the top curve reflects the total decoding time. To put the times for a0 a18 decoding in perspective, the dashed line in the lower plot reflects the total decoding time in a0 a55 decoding. Operations not included in the figures consume so little time that their plots cannot be discerned in the graphs. The times shown are averages of 100 sentences each for length 10, 20, a1a2a1a3a1 , 80.</Paragraph>
    <Paragraph position="2"> IBM Model 4 scores and the BLEU metric. The speed improvements discussed in this paper make multiple randomized searches per sentence feasible, leading to a faster and better decoder for machine translation with IBM Model 4.</Paragraph>
  </Section>
class="xml-element"></Paper>