<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1007">
  <Title>Machine Translation Using Isomorphic UCGs</Title>
  <Section position="6" start_page="34" end_page="34" type="concl">
    <SectionTitle>
5 Current Difficulties
</SectionTitle>
    <Paragraph position="0"> There are several important difficulties that remain unsolved. The first one is how to handle the differences in the freedom of word order in two languages. For instance, Spanish word order is relatively free compared to English. It conveys important stylistic information that should be capturdd in the translation, but which at present gets lost. Another aspect of the same problem is that we would like to be able to recognise all possible word orders in Spanish, without generating them all (as some are intelligible but, sound awkward).</Paragraph>
    <Paragraph position="1"> A possible solution to this could be to include some measure of the degree of &amp;quot;markedness&amp;quot; of a construction in each language. The translation process would attempt to keep the markedness of the two constructions as close as possible to each other. If the grammar specifies that Spanish sentences may be more &amp;quot;marked&amp;quot; than the English, the more marked would never be generated, though they could be analysed.</Paragraph>
    <Paragraph position="2"> Another problem is how the set of basic bilingual signs is to be characterised. That the semantics of SL and TL signs unify is a necessary condition for them to stand in the relation of translation equivalence. It is however insufficient in two ways. First, it must be the case that there is no more specific sign in either language whose semantics unifies with that of the other language, and which is of similar markedness. Secondly, it must be the case that the semantics of the two signs will continue to unify re~. gardless of the derivations into which the signs are incorporated.</Paragraph>
    <Paragraph position="3"> For instance, suppose that the English word leg is associated with the semantics \[leg~of(X,Y)\], and the Spanish word pieraa with \[leg_of(X,Y),human(Y)\]. Although these semantic values do not contradict each other, they will if Y becomes bound to a non-human entity. In this case, the solution is clear deg a further bilino gual sign must be constructed in which English leg is paired with Spanish pats, having the semantics \[leg_of(X,Y),not(human(Y))\].</Paragraph>
    <Paragraph position="4"> Then, either the derivation will eliminate one or the other equiva.lence, or both translations will be produced, which is the desired result.</Paragraph>
    <Paragraph position="5"> It is possible that one monolingual component of a bilingual lexical sign will not be a basic expression in that language. Instead, it must be explicitly constructed in order to be paired with a basic expression in the other language. The unification-based semantics gives an indication of when such a sign-construction process must take place. The flexible categorial approach to the construction of constituents allows the non-standard categories needed to be built.</Paragraph>
    <Paragraph position="6"> in a sense, ell the hard work of this approach takes place at this point. See \[Whitelock 881 for a discussion of the issues involved.</Paragraph>
    <Paragraph position="7"> Finally, ghore is a cluster of problems that impinge on the question of ,:omputatlonal efficiency. It seems unavoidable that certain bilinl~ual signs will need to incorporate either discontinuous or null constituents, or both, ti'om one or the other of the languagesconce ).'J~led.</Paragraph>
    <Paragraph position="8"> g Co~tclusion This paper presents a view of MT that is based on the direct specification of a computable description of a recursive translation relation. We :~mve proposed a system of simultaneous constraints placed on ist,morphie derivation trees in SL and TL whose leaves ,~re elements of a finite set of bilingual signs and whose internal nodes stand in a type-driven compositional relationship to their daughters. \[~ is the combioation of unification and categorial techuiquc.,.~ ~hat makes this idea particularly feasible. The nonst, andard co~,stituents made available in a thll categorial calculus enables iso~lorphic derivation trees to be built; the partiality of the signs aml their combination by unification allows the expression of very precise constraints that both derivations must satisfy. The p,~rtiality of semantic representations is also crucial in determining the set of equivalences - the bilingual lexicon - that form the basis of lhe recursive translation relation.</Paragraph>
    <Paragraph position="9"> There remain many problems with realising this approach in s. practical I~ystem. However, we believe that there are significant advantages to be gained by a direct statement of the translation relation between two languages that is at once declarative, computable and linguistically well-ibunded.</Paragraph>
  </Section>
class="xml-element"></Paper>