<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1602">
  <Title>Coedition to share text revision across languages and improve MT a posteriori</Title>
  <Section position="5" start_page="1" end_page="2" type="concl">
    <SectionTitle>
UNL
</SectionTitle>
    <Paragraph position="0"> (Site on the Web for the Initiation, Information, Validation, Research and Experimentation on UNL [12]) as an experimental basis for our research. It currently allows to:</Paragraph>
    <Section position="1" start_page="2" end_page="2" type="sub_section">
      <SectionTitle>
3.2 Building the lattice-tree correspondence
</SectionTitle>
      <Paragraph position="0"> Let us outline the method (currently under implementation) to compute a &amp;quot;best&amp;quot; correspondence. We start with an MS-L0 lattice linked to the text and a UNL-tree produced in a standard way and linked to the UNL graph. The goal is to establish liaisons between the lattice and the tree, and to order the tree so that it is maximally aligned with the lattice, hence with the text. Suppose we have only an L0-English dictionary.</Paragraph>
      <Paragraph position="1"> First, we enrich the lattice with English lemmas and the UNL-tree with lemmas of L0, producing MS-L0+EN and UNL-tree+L0. Then, we establish links between nodes of the lattice and of the tree having lemmas in common (in L0 or in English), and compute a score for each trajectory in the lattice. The best trajectory is chosen.</Paragraph>
      <Paragraph position="2"> The next phase consists in aligning the tree with that trajectory, using &amp;quot;sure&amp;quot; links as the point of departure, and constraints on the STREE and SNODE liaisons: if there are crossing links, which is possible if two words in the text have similar meanings, preference is given to the link maximizing the proximity in the tree and in the string. Then, liaisons of other types are established:  http://www-clips.imag.fr/geta/User/wang-ju.tsai/ welcome.html lexemes with semantic relations, lexemes with attributes, and MS attributes with attributes.</Paragraph>
    </Section>
    <Section position="2" start_page="2" end_page="2" type="sub_section">
      <SectionTitle>
3.3 Related research
</SectionTitle>
      <Paragraph position="0"> Sending feedback automatically to developers is already done in some MT systems, notably in Taiwan (EKS) and at PAHO [14], but should be much more used than it is. The idea of coedition is also not new: UPM in Madrid uses it to create UNL graphs, Y. Lepage at ATR and Tang E. K. at USM (Penang) have developed editors of string-tree correspondences, Watanabe at IBM-Japan has a very nice interface to edit from a text its underlying dependency structure, the MULTIMETEO system [8] is in effect a coedition system for weather forecasts and their underlying semantic structure, in 6 languages, and there is a project at Xerox working on multilingual generation and free text normalization in restricted domains and typologies (pharmaceutical notices).</Paragraph>
      <Paragraph position="1"> In our case, by contrast, coedition is to happen at the consumer side, not (like at UPM) at the producer side, and there is no specific domain or typology. The idea to derive an abstract semantic tree from an IL representation using alignment techniques and not a rule system embedded in a generator seems also to be new.</Paragraph>
      <Paragraph position="2"> Conclusion Coedition of a natural language text and its representation in some interlingual form seems the best way to share text revision across languages. UNL graphs seem to be the best candidates in this context. We have described an approach where, in the simplest sharing scenario, naive users interact directly with the text in their language (L0), and indirectly with the associated graph. It should also be possible to view and directly manipulate the given UNL graph, a lattice or chart produced by some available free morphosyntactic analyzer, and an abstract tree produced not by analysis, but by a standard transformation from the UNL graph, followed by lexical enrichment in L0, and alignment with the text. When completed, our implementation will make it possible to share revision across languages. We will then have progressed towards merging pivot MT, interactive MT, and multilingual text authoring.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>