<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1604">
  <Title>English-Japanese Example-Based Machine Translation Using Abstract Linguistic Representations</Title>
  <Section position="4" start_page="5" end_page="5" type="metho">
    <SectionTitle>
4 Discussion
</SectionTitle>
    <Paragraph position="0"> It is illustrative to consider some of the factors that contributed to these results. Table 2 shows the number of transfers per sentence and the number of LF nodes per transfer in versions of the system evaluated in October 2001 and May 2002. Not only is the MSR-MT finding more LF segments in the Mindnet, crucially the number of nodes transferred has also grown.</Paragraph>
    <Paragraph position="1"> An average of two connected nodes are now transferred with each LF segment, indicating that the system is increasingly learning its translation knowledge in terms of complex structures rather than simple lexical correspondences.</Paragraph>
    <Paragraph position="2"> It has been our experience that the greater MSR-MT's reliance on the Mindnet, the better the quality of its output. Table 2 shows the sources of selected word classes in the two systems. Over time, reliance on the Mindnet has increased overall, while reliance on dictionary lookup has now diminished to the point where, in the case of content words, it should be possible to discard the handcrafted dictionary altogether and draw exclusively on the contextualized resources of the Mindnet and statistically-generated lexical data. Also striking in Table 2 is the gain shown in preposition handling: a majority of English prepositions are now being transferred only in the context of LF structures found in the Mindnet.</Paragraph>
    <Paragraph position="3"> The important observation underlying the gains shown in these tables is that they have primarily been obtained either as the result of LF improvements in English or Japanese (i.e., from better sentence analysis or LF construction), or as a result of generic improvements to the algorithms that map between LF segments (notably better coindexation and improved learning of mappings involving lexical attributes). In the latter case, although certain modifications may have been driven by phenomena observed between Japanese and English, the heuristics apply across all seven languages on which our group is currently working. Adaptation to the case of Japanese-English MT usually takes the form of loosening rather than tightening of constraints.</Paragraph>
  </Section>
class="xml-element"></Paper>