<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1123">
  <Title>Empirical Lower Bounds on the Complexity of Translational Equivalence [?]</Title>
  <Section position="7" start_page="983" end_page="983" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> This paper presented evidence of phenomena that can lead to complex patterns of translational equivalence in bitexts of any language pair. There were surprisingly many examples of such patterns that could not be analyzed using binary-branching structures without discontinuities. Regardless of the languages involved, the translational equivalence relations in most real bitexts of non-trivial size cannot be generated by an inversion transduction grammar. The low coverage rates without gaps under the constraints of independently generated monolingual parse trees might be the main reason why &amp;quot;syntactic&amp;quot; constraints have not yet increased the accuracy of SMT systems. Allowing a single gap in bilingual phrases or other types of constituent can improve coverage dramatically.</Paragraph>
  </Section>
class="xml-element"></Paper>