<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-0910">
  <Title>Ivan.P.Bretan@telia.se Robert.H.Eklund@telia.se Mats.G.Wiren@telia.se</Title>
  <Section position="2" start_page="0" end_page="55" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The basic idea of this paper is simple and uncontroversial. All natural languages are in some sense similar (some are obviously very similar), so software written to process one language ought to some extent to be applicable to other languages. If the languages L1 and 52 are similar enough, then it should be easier to recycle software applicable to L, than to rewrite it from scratch for L2.</Paragraph>
    <Paragraph position="1"> This paper describes two related approaches in this general direction, which have been successfully applied within the Spoken Language Translator (SLT) project (Rayner and Carter, 1997). The first is the most obvious: we start with a functioning grammar and lexicon for L1, and port it to the similar language L2. This is not, of course, a novel idea, but we think that we have refined it in a number of ways. In particular, we show that it is practically feasible in the case of sufficiently close languages to generalize an existing grammar for L1 to cover both L1 and L2 (i.e. produce a single grammar which through the setting of a single parameter becomes valid for either language). We also describe a method which makes it possible to port the language-dependent lexicon for L1 so as to maximize sharing of data between the systems for the two languages.</Paragraph>
    <Paragraph position="2"> The second idea is specifically related to translation. Suppose we have already developed sets of transfer rules for the two language-pairs L1 --+ L2 and L2 ~ L3. We describe a method which allows us to compose the two sets of rules off-line to create a new set for the pair L1 --+ L3.</Paragraph>
    <Paragraph position="3"> Both methods might be said to operate according to the principle memorably described by Mary Poppins as &amp;quot;Well begun is half done&amp;quot;. They do not solve either problem completely, but automatically take care of most of the drudgery before any human has to become involved. In each case, the initial result is a machine-written set of linguistic data (lexicon entries and transfer rules) which is not quite adequate as it stands; a system expert can however clean it up into satisfactory shape in a small fraction of the time that would have been required to write the relevant rules and lexicon entries from scratch.</Paragraph>
    <Paragraph position="4"> The practical experiments we describe have been carried out using versions of the SLT system involving the languages English, French, Swedish and Danish. Initial results are extremely promising. In particular, we were able to combine both methods to create fairly credible Swedish-to-French and Englishto-Danish spoken language translation systems I us~In fact we do not currently use a Danish speech synthesizer, but it would be straightforward to incorporate  ing only a few person-weeks of expert effort.</Paragraph>
    <Paragraph position="5"> The rest of the paper is structured as follows. Section 2 gives a very brief overview of the relevant aspeers of the SLT system. Section 3 describes the methods we have developed for porting linguistic descriptions between closely related languages. Section 4 summarizes the transfer composition method.</Paragraph>
    <Paragraph position="6"> Section 5 describes preliminary experiments.</Paragraph>
    <Paragraph position="7"> 2 An overview of the SLT system The SLT system has been described in detail elsewhere (most recently (Rayner and Bouillon, 1995; Rayner and Carter, 1997)), so this section will only provide the minimum information necessary to understand the main body of the paper.</Paragraph>
    <Paragraph position="8"> The language-processing (translation) part of the system is supplied with N-best sentence hypotheses by the system's recognizer, and itself uses a hybrid architecture, which combines rules and trainable statistical models. To summarize the argument from (Rayner and Bouillon, 1995), there are good reasons for requiring both these components to be present.</Paragraph>
    <Paragraph position="9"> Rules are useful for capturing many kinds of regular linguistic facts that are independent of any particular domain of application, prime examples being grammatical agreement and question-formation. In contrast, there are other types of phenomena which intuitively are more naturally conceptualized as idiosyncratic and domain-dependent: the most obviotis examples here are word-choice problems.</Paragraph>
    <Paragraph position="10"> The system uses two translation mechanisms, applied bottom-up in parallel (Rayner and Carter, 1997). The primary, rule-based translation mechanism performs transfer at the level of Quasi-Logical Form (QLF), a type of predicate/argument style notation (Alshawi et al., 1991). The source- and target-language grammars provide a declarative definition of a many-to-many mapping between surface form and QLF. The grammars are domainindependent, and can be compiled to run efficiently either in the direction surface form ~ QLF (analysis) or QLF --+ surface form (generation). In transfer, unification-based rules are used to define a space of possible candidate translations; domaindependent, statistically trained preferences then select the most preferred candidate translation. This division of effort has the important consequence of allowing the transfer rules to be fairly simple, since much of the complexity is &amp;quot;factored out&amp;quot; into the trained preferences.</Paragraph>
    <Paragraph position="11"> In order to deal with the brittleness inherent in az~y rule-based system, a second, much less sophistione. null cated translation method is also used, which simply maps surface phrases from the source language into possible target-language counterparts. We refer to the backup method as &amp;quot;word-to-word&amp;quot; (WW) translation. The two methods are combined, roughly speaking, by using rule-based QLF transfer to translate as much as possible, filling in any gaps with applications of the WW rules.</Paragraph>
    <Paragraph position="12"> The parts of the system of central interest here are the rule-based components, in particular the morphologies, grammars, lexica and transfer rules.</Paragraph>
    <Paragraph position="13"> Morphologies are written in a variant of two-level morphology (Carter, 1995), and grammars in a unification-based formalism (Alshawi (ed), 1992).</Paragraph>
    <Paragraph position="14"> The lexicon for each language is divided into three main parts: Domain-independent function (closed class) word entries are written directly in terms of definitions of suitable feature-value assignments, and can from a software-engineering standpoint be regarded as part of the grammar.</Paragraph>
    <Paragraph position="15"> A collection of language-dependent but domain-independent macros define the feature-value assignments needed for each type of regular content-word, e.g. &amp;quot;count noun&amp;quot;, &amp;quot;transitive verb&amp;quot; and so on. These macros are called paradigm macros.</Paragraph>
    <Paragraph position="16"> Content word entries, which in general may be domain-dependent, are defined in terms of these lexical macros. An entry of this kind contains the following information: the name of the relevant macro, the base surface form of the word, the associated logical-form (QLF) constant, and if necessary a pointer to the correct inflectional type (conjugation or declension).</Paragraph>
    <Paragraph position="17"> Structurally, transfer rules have much in common with lexicon entries. (Bear in mind that conventional bilingual and monolingual dictionaries have similar structures too). A small set of domain-independent transfer rules encode cross-linguistic divergences significant enough that they need to be represented at the QLF level: these rules may contain arbitrary pieces of QLF form. The majority of the transfer rules, however, are &amp;quot;atomicatomic&amp;quot; : they associate a logical-form constant from the source language with one or more logical-form constants from the target language. Transfer rules of this type have a close connection with the macrodefined monolingual content-word lexicon, and may also be domain-dependent.</Paragraph>
  </Section>
class="xml-element"></Paper>