<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-3003">
  <Title>Can the Internet help improve Machine Translation?</Title>
  <Section position="9" start_page="221" end_page="221" type="concl">
    <SectionTitle>
8 Technical Challenges and Future Work
</SectionTitle>
    <Paragraph position="0"> The Rule Refinement process is not invariable. It depends on the order in which refinement operations are applied. In batch mode, the RR module can rank Correction Instances (CI) in such a way as to maximize translation accuracy. Suppose that the first CI (CI1) triggers a bifurcation of a grammar rule, like the one we see in the example described in Section 5. After that, any CI that affects the same rule that got bifurcated, will only modify the original rule (NP,8) and not the copy (NP,8').</Paragraph>
    <Paragraph position="1"> If the constraint that enforces determiner-noun agreement were missing from the original rule, say, the copy (NP,8') would not have that constraint added to it, and so another example with the pre-nominal adjective exhibiting that agreement error would be required (CI2: *Irina es un gran amiga), before the system added the relevant constraint to NP,8'. However, if we can detect such rule dependencies before the refinement process, then we can try to find an optimal ranking, given the current set of CIs, which should result in higher translation accuracy, as measured on a test set.</Paragraph>
    <Paragraph position="2"> Another interesting future direction is enhancing the Rule Refinement system to allow for further user interaction. In an interactive mode, the system can use Active Learning to produce minimal pairs to further investigate which refinement operations are more robust, treating the bilingual speaker as an oracle. We hope to explore the space between batch mode and a fully interactive system to discover the optimal setting which allows the system to only ask the user for further interaction when it cannot determine the appropriate refinement operation or when it would be impossible to correctly refine the grammar and the lexicon automatically.</Paragraph>
  </Section>
class="xml-element"></Paper>