<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1007">
  <Title>Machine Translation Using Isomorphic UCGs</Title>
  <Section position="5" start_page="0" end_page="34" type="evalu">
    <SectionTitle>
4 UCG and Isomorphic Grammars
</SectionTitle>
    <Paragraph position="0"> The principle of Isomorphic Grammars is realised in UCG by means of bilingual signs. Bilingual rules, which combine bilingual signs, may be defined in terms of how monolingual rules combine the monolinguM parts of the sign.</Paragraph>
    <Paragraph position="1"> As was mentioned, monolinguM UCG signs consist of four features: Phonology, Syntax, Semantics, and Order. A bilingual sign is merely a sign with top-level attributes source and target having monolingual signs as their values, and in which source se~ mantics and target semantics share their value. Since transla~ ties must preserve semantics, this sharing of values is a necessary condition. In the general case, however, it is not sufficient (sev section 5).</Paragraph>
    <Paragraph position="2"> The Bilingual sign can easily be decomposed into, or built up from, a Source sign and a Target sign (having a common Semantics), by a Prolog predicate decompose(Bllingual_Si~t, Source_Sign, Target_Sign).</Paragraph>
    <Paragraph position="3"> Combination of two monolingual signs is defined by two predicates: null  source_combine(S1, S2. S).</Paragraph>
    <Paragraph position="4"> target_combine(Tl, T2, T).</Paragraph>
    <Paragraph position="5"> which combine their first two arguments to give the third. The crucial difference between these two predicates is as follows: source_combine requires that the order feature of S1 and $2 is consistent with the phonology of S, while target_combine ensures that the phonology of T is consistent with the order of T1 and T2. This enables differences in word order in the Source and Target Languages to be accounted for, as shown below. The two monolingual modes of combination above are used to define bilingual combination through a predicate:  bilingual_combine(B1, B2, B):decompose(B1, S1, T1), decompose(B2, S2, T2), source_combine(S1, $2, S).</Paragraph>
    <Paragraph position="6"> target_combine(Tl, T2, T), decompose(B, S, T).</Paragraph>
    <Paragraph position="7">  The way in which differences in word order are dealt with may be illustrated by the translatioin equivalence between an adjective-noun combination in English and a noun-adjective combination in Spanish. For the sake of simplicity, only the features for phonology, syntax and order are included.</Paragraph>
    <Paragraph position="8"> The predicate source_comblne allows two combinations:  (I) gz A/B s: C ----* ptm AWl W2 o: pro (2) s: C p: W2 p~ WI W2 o: post s: A/B ~ m A  (where the active part of the functor sign unifies with the argument sign) The predicate target_combine, on the other hand, allows the above two combinations, and in addition the two order-reversing ones:  s: A/B m C ~ p: W2 Wl o: post s: A (4) e: C p: W2 p: W2 Wl s: A/B --4 .: A o: pre  Let us then examine how the English expression red book gets translated into the Spanish llbro rojo, in which the order of the adjective and noun are reversed.</Paragraph>
    <Paragraph position="9"> The bilingual signs are: \[ sre:p: red tgtsp~ tgtzs: noun/ s:z post and \[ src:p: book 1 tgt:p: nbro grc:9~ noun tgt:s: noun These will get decomposed into their source and target constituents, which may only be combined using (1) and (3) above, respectively:  st noun p: book (1} p: red book m noun m noun a~ noun o: pre Currently, we assume the existence of four bilingual signs core responding to the English word red, since the Spanish adjective has four combinations of gender and number. Only that sign re~. resenting the contextually correct translation equivalence will be incorporated in the derivation. In a practical system, there would be a single bilingual sign whose Spanish component has disjunctive (or unspecified) values for gender and number, and the incorporation of this sign into the derivation will eliminate the disjunction (or bind the variables).</Paragraph>
    <Paragraph position="10"> Unlike Landsbergen's approach, it is not necessary to specit~y that the rules which combine the SL and TL expressions must be the same. Because of the type-driven mapping between syntax and semantics, if two pairs of signs stand in the translation relation, then so will the pair of signs resulting from their combination, regardless of the rule used.</Paragraph>
  </Section>
class="xml-element"></Paper>