<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1070">
  <Title>Lieven Oaspaert</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
LINGUISTIC DEVELOPNENTB IN EUROTRA SINCE 1983.
</SectionTitle>
    <Paragraph position="0"> I wish to put the theory and metatheory currently adopted in the Eurotra project (ArnoO6) into a historical perspective, indicating where and why changes to its basic design for a transfer-based MT (TBMT) system have been made.</Paragraph>
    <Paragraph position="1"> Let T. be some theory of representation, inducing sets of representations R~ and Rt for languages L,, and L~ (seen as sets of texts), respectively.</Paragraph>
    <Paragraph position="2"> Transfer-based translation is described as follows:</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="296" type="metho">
    <SectionTitle>
AN
TRF
</SectionTitle>
    <Paragraph position="0"/>
    <Paragraph position="2"> where AN, BEN and TRF are binary relations, and TRA is the composition of AN, TRF and GEN, i.e.</Paragraph>
    <Paragraph position="3"> interpretation was to be expressed as an orbered tree with complex property lists on the nodes, which was manipulated by two basic operations, viz. tree transformations and lexical substitution. The BETA preoccupation with robustness, on the other hand, made them require that all linguistic information about texts should be merged into one single gds. On failure to compute parts of a deeper linguistic dimension, the intuition went, some clever algorithm could be used to extract from the gds an equivalent piece of representation on the next less pretentious dimension. The logical extreme of this reasoning was that, if all else failed, it should be possible to recover the original text from the gds.</Paragraph>
    <Paragraph position="4"> Grenoble, however, had perceived the usefulness of dependency theory (DT) for TBNT. &amp;quot;\[here is a sense in which DT is a lexically oriented theory of language, and, in the end, translation is a question of getting the right translation for words.</Paragraph>
    <Paragraph position="5"> Nevertheless, the marriage between DT and the gds design led to (I) procrustinated linguistics, and  (2) a formalism with untractable semantics.</Paragraph>
    <Paragraph position="6"> (i) AN ~ L~ x R,~, GEN ~ R~ x L~, TRF s R~ x R~ (ii) TRA = AN o TRF o GEN  We also need to introduce two parameters, viz.</Paragraph>
    <Paragraph position="7"> hypotheses about T .... A theory (e.g. for the AN relation) is m ultistratal when It consists of a set of subtheories {tl,t2,...,tn}, each characterising a set of representation R,, such that i.i. ~Io~ostratal.</Paragraph>
    <Paragraph position="8"> The advocated representation theory was not stratified in any interesting sense. Rather, the whole burden of modularising the relation between text and representation was put on the translation of the relation into a procedure: discussions about never brought to hear.</Paragraph>
    <Paragraph position="10"> Otherwise, a theory is monostratal, A theory T is mult!_dim_ebsional when descriptions of linguistic objects along several linguistic dimensions are merged into one single representational mbject. The notion of linguistic dimension is meant to correspond to some organising principle for a theory of representation (e.g.</Paragraph>
    <Paragraph position="11"> constituency, grammatical relations, logical semantics, ate.l. Otherwise, a theory is monodimensional.</Paragraph>
    <Paragraph position="12"> In what follows we describe the various Eurotra approaches to TBMT in terms of this basi~ model, Initially, due to its BETA inheritance, Eurotra adhered to a monostratal multidimensional model for TBMT. Computationally, it was based on the Grenoble formalism of the g~n~E~ateu_r d~._stEuctur~s (g.ds). Linguistically, it advocated a diluted form of dependency theory as a basis for TBMT.</Paragraph>
    <Paragraph position="13"> The observation that theoretical linguistics had been incapable of providing a practically applicable basis for translation had led GrenoOle to build almost no linguistic commitment into the gds formalism. Every possible form of linguistic 294.</Paragraph>
    <Paragraph position="14"> \]he innovation of \[Arno83\] was its attempt to derive requirements on T. from a set of more abstract principles, seen as a theory of MT providing a framework within which possible substantive theories for TBMT could be devised and compared. The weakness of the framework was to seek to motivate the tools inherited from BETA a-posteriori. Its merit was to be a partial theory of TBMT, independent of the inheritance.</Paragraph>
    <Paragraph position="15"> Its major concern was directed at elucidating the division of labour between AN, TRF and OEN, and at deriving implications on T. from this understanding. The pivotal principles of the framework that have survived the many face lifts of the Eurotra model are i soduidy_ and Q-d~E~D~.g!2.</Paragraph>
    <Paragraph position="16"> The principle of isoduiy allowed for a principled definition, in terms of properties of T,., of the domain of the GEN relation of some language in terms of the nodomain of the AN relation for that same language, thus indirectly defining .the TRF relation. The principle of O-differentiation required that T. should be sufficiently expressive to ensure that all meaning aspects of text that are relevant for translation (called 'Q') be represented in members of R. The two principles together&amp;quot; provided a basis for designing a transfer device that was (i) developmentally simple, and (2) Q-preserving. These are necessary features of any multilingual TBMT system striving for good-quality translation.</Paragraph>
    <Paragraph position="17"> 1.2. ~lultidimensior, al.</Paragraph>
    <Paragraph position="18"> Despite its success in providing an initial framework for Eurotra, lame83\] failed dismally when it came to deriving from it a substantive linguistic representation theory. The failure was not unrelated to the absence of motivation for the GETA vestiges. The gds comprised a flat geometry and a rich decoration on the nodes. Given the requirement of merging, th~ geometry for all dimensions (text string, morphology, surface syntax, deep syntax, semantics) had to be very similar: this was only possible by making the geometry quite meaningless, and by putting the whole expressive burden on the labelling of nodes. The need to preserve surface word order (robustness) gave geometry its only interesting task: the representation of word order through the ordering of sister nodes, Within a merged approach, this requirement led to the arbitrary interdependence of the subtheorios for the various linguistic dimensions. The problem was most tangible in the design of a subtheory of \]',, for a semantic dimension. T~, became unnecessarily complex and inconsistent. Given the absence of linguistic ~ommitments built into the tools and the failure of the framework to answer substantive linguistic questions, debates about the relative merits of particular representational choices wore inconclusive.</Paragraph>
    <Paragraph position="19"> We give an example of linguistic procrustination, Surface word order being represented by the order of sister nodes in the merged tree (the gds), tree geometry was seen as ordered. Tins geometry of dependency representations, on the ether hand, are normally unordered, Tile way out was a refashioning of DT as a compromise between DT and X-theory with a single bar: a subset of the information about the governing node was lowered into the mdbtree representing its dependents and to require that the subtree he ordered conforming to the position o~ elements in the input text, lbi~ worked badly with all sorts of difficult linguistic phenomena: exocentri(: constructions (e.g. ~onjuection), gapoing, discontinuity, long-distance dependencies, etc. Much of the linguistic research, then, was aimed at overcoming these problems in a principled way by means of a theory of empty elements. Although the latter was intuitively consistent, it caused such an increase in the complexity of the formalism that the latter defied any coherent formal characterisation.</Paragraph>
    <Paragraph position="20"> The first design was, amongst other things, unable to flesh out the problem of robustness, Combining a muitidimensienal representation with a b~sically all paths combinatorial algoritbm led to the inability to rely on the actual computation of combinations of information required by the safety net algorithm.</Paragraph>
    <Paragraph position="21"> The second deuign (which was never formally accepted by the project) purported to solve this problem, without eliminating multidimeneion-alitv. It was multistratal and multidimensional.</Paragraph>
    <Paragraph position="22"> 2.1. Rultistr~tal.</Paragraph>
    <Paragraph position="23"> It was observed that the representations induced by T,, had to meet two (possibly conflicting) requirements: (i) they had to have sufficient expressive power to allow for adequate translation via simple transfer, and (2) their computation had to be feasible. As a consequence, T. was split into two subtheories, T~ and T~, were the former was directed at the needs of adequacy for simple transfer and the latter to the reliability of presence of a consistent representation from which either the more pretentious T~ representation was reached or, alternatively, translation via less.simple transfer was possible. The model that emerged was the following:  The motivation .\[or this design hinged on (i) the fact that the f-stratum could make use of know-how in computational linguistics, (2) the f-stratum was a good starting point for innovative research on what T~. stlould be for multilingoal TBMT, (3) the model gave content to the notion of safety nets (rmbostness), (4) developmental issues, The claim made was that with a monolithic l., the formulation of safety nets is hindered hy the hybridity prohlem~ their' input domain coold be any unpredictable mombination of feasibly computable and adequate information on several dimensions in the gds. \]he new design provided the f-stratum as a more reliable basis for' safe safety nets.</Paragraph>
    <Paragraph position="24"> 2.2. Hultidi~nsio~aI.</Paragraph>
    <Paragraph position="25"> This feature of the design did not change, Instead of one multidimensional representation, we now had two, No further attempt was mmde, however, to justify the use of multidimensional representations. Given the rejection of theoretical modularity on the basis of considerations of reliability of computation, the only course to take seemed to be to abandon the multidimensional view itself and to let the strata themselves v, epresenh linguistic dimensions. The new model became multistratal and monodimensional, 3.1. khlltistratal.</Paragraph>
    <Paragraph position="26"> T,, was described as a sot of iodependently defined subtheories for&amp;quot; representing normalised text (no), morphology (me), surface syntax (ss), deep syntax (dm) and semantics (sem), They were concogtumlly related to each other, however, by being based on a common central notion of dependency defined in terms of slotfilling and modification, A strength of this move is that linguistics in Eurotra could now profit from linguistic work in the outside world, The proposal suffered, l'lowever~ from the absence of a clear view on what sorts of dedicated operations were needed to actually map between arbitrarily different dependency trees. Nor were considerations of the computational complexity of arbitrary treetransformation formalisms taken into account in the definition of the levelo, A proposal to relate all these levels to each other by giving them all a lexicalist underpinning was rejected by the C,E.C.</Paragraph>
    <Paragraph position="27"> Finally, a stratificatienal strategy was imposed on the makers of the design, with the (unjustified) intuition that it would provide a basis for the incorporation of safety nets into the model.</Paragraph>
    <Paragraph position="28">  The model now roughly looked as follows (with question marks indicating undefined parts):  dimension: the gds approach was completely abandoned.</Paragraph>
    <Paragraph position="29"> The theories identified described the representation of normalised text strings, the internal structure of words, the surface dependency, the canonical dependency and the semantic dependency of the input texts.</Paragraph>
    <Paragraph position="30"> 4. The #resent desi~_LCde~T85jA\[!oB6\]+-The properties of the current Eurotra design constitute the topic of Arnold &amp; des Tombe's paper in this volume. Here, I merely relate it to previous hypotheses about the Eurotra translation model, The design is multistrata\] and monodimensional and can be depicted as follows:  Each stratum corresponds to an autonomous generating device for a representation language, Each generator consists of a set of atoms and a set of constructors that together allow for the generation of L(8), a set of formally well-formed derivation trees. The latter are then evaluated (by unification) to a set of meaningfull representations, R(G).</Paragraph>
    <Paragraph position="31"> The intuition underlying this model is that translation between natural language texts can be split up into a sequence of more IE~Y~_ translations between elements of adjacent generators. Adjacent generators must be devised so that the primitive translations that obtain are also simll_e, This is taken to mean that primitive translations must be (i) compositional and (2) oneshot. The justification for compositionality is the intuition that the translation of some expression E  is a straightforward function of the translation of E's parts and of the way these parts ere put together, The latter is required to restrain the complexity of this function: the codomain of a primitive translation must always be well-formed in terms of the target generator, This forbids internal strategy inside translators.</Paragraph>
    <Paragraph position="32"> The project is examining various hypotheses about particular instantiations of this core model: e.g. translators could perform any one of the following four mappings: (i} derivation to derivation, (ii) derivation to representation, (iii) representation to derivation and (iv) representation to representation. Possibility (i) was found to be too restrictive. We now study possibility (iii). Note the similarity between (iv) and the structural correspondence approach adopted in LFG for mapping between information structures of a different nature.</Paragraph>
    <Paragraph position="33"> 4.2. Monodi~ensional.</Paragraph>
    <Paragraph position="34"> The current strata envisaged are normalised text, morphology, configurational surface syntax, relational surface syntax and semantics. Morphology is based on work on word grammar as independent of phrase structure grammar. Configurational syntax draws from the X-theory literature. Relational syntax representations resemble LFG f-structures.</Paragraph>
    <Paragraph position="35"> The semantic stratum, finally, is not yet fully specified: this has to do with the very special requirements that translation by means o~ simple transfer puts on e semantic representation theory. The point is, however, that the non-semantic levels are claimed to be feasible (cfr. f-stratum in 2) and that they can thus provide a basis for researching a translation-oriented semantic theory,</Paragraph>
  </Section>
  <Section position="3" start_page="296" end_page="296" type="metho">
    <SectionTitle>
5. Conclusion.
</SectionTitle>
    <Paragraph position="0"> I hope to have slightly lifted the veil that has hidden the Eurotra project from the scientific community for a number of years. It has become clear, hopefully, that the Eurotra design has become more homogeneous and that it constitutes a valuable step towards a better understanding of the problem of machine translation.</Paragraph>
  </Section>
class="xml-element"></Paper>