<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1006">
  <Title>MACHINE TRANSLATION IN EUROPE</Title>
  <Section position="3" start_page="0" end_page="30" type="metho">
    <SectionTitle>
WHAT IS SPECIFIC TO MACHINE
TRANSLATION
</SectionTitle>
    <Paragraph position="0"> Bar-Hillel (1964:212), in one of his numerous papers on translation, identified the following five prerequisites for high quality human translation:  (1) competent mastery of the source language, (2) competent mastery of the target language, (3) good general background knowledge, (4) expertness in the field, (5) intelligence (know-how).</Paragraph>
    <Paragraph position="1">  For MT, the first four are obviously necessary and moreover, remain research topics for NLP in general. The last point could simply be replaced by 'the ability to establish a correspondence between the source and target language.' For humans, it is well-known that bilinguals are not necessarily good translators, and discussing intelligence in the context of MT (given the lack of any theory of translation and the current relatively undeveloped state of the art) would only lead to philosophical speculation. Perspectives on combining Artificial Intelligence (A.I) and MT can be found in the numerous discussions by Wilks (e.g., Wilks, 1973), mentioned here in view of his long history in working on this topic both in the US and in Europe.</Paragraph>
    <Paragraph position="2"> The major portion of work in any MT project is concentrated on points (1) and (2) and hence does not differ from any other NLP project except in so far as it involves descriptions of two languages. Limiting the application to a well defined semantic domain or specific corpus reflects the concern for points (3) and (4). One proposal addressing (4) explicitly is to include expert system modules in addition to linguistic rules (Boitet &amp; Gerber, 1984) or to divide the system into expert tasks (Johnson &amp; Whitelock, 1985). In this perspective, MT can thus be viewed as a sub-field of NLP which in some cases, also means incorporating general AI techniques.</Paragraph>
    <Paragraph position="3"> A number of topics specific to MT can be identified, though some of them may be shared by a portion of the CL community, albeit for different reasons. For example, the adequate level(s) of representation, will be determined in MT by the ability to express the relevant facts about two natural languages and to state the translation relation between them; similarly, in a database query system, the concern is to map between two representations where one is an abstraction of the natural language expression but the other is an expression in a formal query language (which has already been def'med).</Paragraph>
    <Paragraph position="4"> Current issues which often figure in the literature and which we will look at in somewhat more detail are:  natural language phenomena) This list is by no means exhaustive, but every MT researcher will have at least one view on most of these topics. Whereas in monolingual lexicography for computational applications, progress has been made on at least a subset of the basic information necessary, this is not the case in the field of bilingual lexicography (Warwick, 1987). In dictionary publishing houses (of interest as a starting point for building machine tractable dictionaries) bilingual lexicographers are seen as the poor cousins, and across any two dictionaries there is essentially no consensus on which translations should be included nor on how to code them. Each project or product has a notation so specific to the system in which it is used, that it is of little use to others. It is notable that almost all work on using machine readable dictionaries has concentrated on monolingual dictionaries, and for the most part only on English; the two European languages which have received some attention are Italian (by the center in Pisa) and German (by the Institute for the German language, IDS, Mannbeim, and the IKP at the University of Bonn).</Paragraph>
    <Paragraph position="5"> In light of this situation and given the crucial role that the lexicon plays in any system, this topic has been identified as a field in its own right. Two European initiatives, EUROTRA-7 and MULTILEX, are currently underway under the title of 'reusability of lexical resources' (McNaught, 1990). The EUROTRA-7 project (ET-7, 1990) is currently conducting a broad survey on what resources are available with a view to developing standards for mono-, bi-, and multi-lingual lexical descriptions. 3 Another project arising from work in EUROTRA, is the EUROTRA-Lexic project whose aim is to build one bilingual dictionary adequate for human and machine needs, involving partners from the publishing, private and academic sector (Eurotra-Lexic, 1989). MULTILEX is an ESPRIT project whose aims are to develop a European multilingual and multifunctional lexicon which will be tested in prototypical applications. The Acquilex project, also funded under ESPRIT, is another project working on standards and a prototype implementation for acquiring and coding mono- and bilingual lexical information. Bilingual concordancing is a relatively new topic which will most likely grow in importance, in parallel with the current trend to make use of large corpora. Instead of building an MT system based solely on a corpus and its translation, as reported on in Brown, et al. (1988), 4 emphasis in Europe has been on developing tools to navigate through texts as an aid for the linguist or lexicographer working on the problem of characterizing the context for the translation of words (cf. Picchi &amp; Calzolari, 1986 and Warwick, et al., 1990). In Czechoslovakia, at the Charles University, Prague, a project has just begun to build a bilingual concordance environment for Czech and English as a first step in work on a new machine translation project (personal communication). A somewhat related project, which plans to make use of structured texts and their translations is the BSO Bilingual Knowledge Bank system (Sadler &amp; Vendelrnans, 1990). Projects in MT using statistical methods are just beginning in a number of centers, but information about these activities has not yet reached the public domain. 5 The result of this survey will be made available to the general public. Contact: Mr. R. Cenconi, CEC, B410Os, Jean Monnet Bldg., 2920 Luxembourg.</Paragraph>
    <Paragraph position="6"> Other European IBM centers may also begin projects with similar methods and goals (personal communication).</Paragraph>
    <Paragraph position="7"> Two centers which have reported on such plans are the University of Stuttgart where work on alignment of texts, in view of linking not only words but also phrases, is planned, and the Rosetta project at Philips, which is considering collecting knowledge of the world for a separate semantic component by statistical methods (personal communication).</Paragraph>
    <Paragraph position="8">  Representation issues have been at the core of machine translation since its very beginning. One of the current trends is to use existing linguistic theories as the basis for translation. One project in Stuttgart has been exploring the use of an extended LFG representation as the basis for translation (of, Kaplan et al. (1989), also Sadler et al., 1990) whereas a project in Berlin, has taken GPSG syntactic representations as the starting point for adding additional levels (Busemarm &amp; Hauenschild, 1988). 6 Other examples of basing MT on state of the art CL methods and theories include the use of situation schemata as the basis for translation Rupp (1989), defining transfer over 'quasi logical forms' developed at SRI, Cambridge (Alshawi et al., 1991) and derivation trees of a Montague style (Landsbergen, 1987 &amp; Appelo, et al., 1989). The Eurotra MT project bases its work on a number of levels essentially corresponding to traditional linguistic categories, i.e., morphology, syntax, and semantic relations plus a special level for transfer known as the interface structure (Arnold, et al., 1985). These are explicit levels and a transfer mechanism allows a mapping between them (Arnold, 1987, Bech&amp; Nygaard, 1988). 7 The search for the ideal level of representation for expressing a translation relation raises the well known issue of transfer vs. interlingua. Almost all projects currently underway in Europe essentially rely on two independent levels, one per language, with an explicit mapping between the two. One exception is the DLT project which uses Esperanto as a pseudo interlingua (Witkam, 1988); a choice of representation which has &amp;quot;aroused a lot of skepticism&amp;quot; in the community (op cit., p.756). The one project theoretically committed to an interlingua is the Rosetta system, a project noteworthy for its theoretical commitment and steady development over the past ten years (Landsbergen, 1989). The work on multilingual generation based on conceptual hierarchies in the project Polygloss (Emele et al., 1990) may also be considered a type of interlingua system.</Paragraph>
    <Paragraph position="9"> A very practical reason for the popularity of the transfer model, especially for systems to treat more than one language pair, stems from the inherent difficulty of defining a representation adequate for more than one natural language, especially when the competent people who might work on such an interlingua cannot work together in one place (as in the Eurotra project which is spread all over Europe). To define an interlingua for translation requires expertise in linguistics (applied to the languages in question), plus a large range of issues often labelled 'extra-linguistic,' knowledge of translation (to ensure that a mapping is possible) as well as familiarity with formalisms for representing the information (as found in much of the work in AI).</Paragraph>
    <Paragraph position="10"> The German government has demonstrated an important commitment to the field of MT as well as NLP in the numerous projects and positions it has financed. The two projects are supported by the government under a program known as Eumtra related research projects (Eurotra Begleifforschung).</Paragraph>
    <Paragraph position="11"> The original motivation for separate levels in the monolingual portion came in part from practical concerns that for each language some progress could be measured (e.g., each language had accounted for molphology, syntax, etc.). For the bilingual portion, this choice had a theoretical motivation, i.e., transfer as the model for translation, and the mechanism for mapping between levels was extended to distinct monolingual levels.</Paragraph>
    <Paragraph position="12"> Two topics which have gained importance as a result of adopting a transfer model are the formalization of the transfer component and the issue of reversibility of grammars (cf.</Paragraph>
    <Paragraph position="13"> Isabelle, 1988). Reversibility of a grammatical description ensures that all necessary information has been accounted for in the representation for both parsing and generation and also defines what the output of ~ansfer from one language to another must be. If the representation from analysis is underspecified, in a reversible description this will become apparent as overgeneration. For translation, the notion of reversibility helps to test whether the relation is symmetric, an attractive working hypothesis for a theory of translation.</Paragraph>
    <Paragraph position="14"> The transfer mechanism in earlier work, as in older systems such as the different versions of ARIANE (Boitet, 1988), a system developed in Grenoble at one of the oldest European centers for machine translation; METAL (White, 1985), the Siemens German-English system initially developed in the States; 8 and the SUSY system (Maas, 1988), a system developed in the 1970s in Saarbruecken, was an arbitrary tree-to-tree transformation. Compositionality and declarativity have since become basic tenets as a means of overcoming the ad hoc and procedural mechanisms those systems employed (cL Landsbergen, et al., 1989 on compesitionality in the Rosetta system and des Tombe, et al., 1985 for a discussion of &amp;quot;relaxed compositionality&amp;quot; in the Eurotra framework). Current work concentrates on constraining the transfer mechanism within a well defined computational model, e.g., transfer rules are defined between feature structures and the mechanism is based on unification (cL van Noord, et al., 1990, Russell et al., 1991 and Wedekind, 1988). Unification as the basis for MT systems serves as a basis for quite a number of MT projects in Europe (in addition to the above mentioned el. also Carlson &amp; Vilkuna, 1990 and Zajac, 1989); its advantages are the well understood formal aspects and the declarative nature of the rule schema for a given implementation.</Paragraph>
    <Paragraph position="15"> As a conclusion to this section, let me mention a few outstanding topics that are under investigation in the CL literature but are noticeably absent in the MT literature in Europe. The approach to translation as more of an AI problem (cL numerous papers by Nirenburg and Wilks) has not received much attention in Europe. This is perhaps due to the fact that MT projects are more often found in (computational) linguistic departments than computer science and AI labs. The generation work in MT has mainly concerned itself with reversibility issues and has hardly taken into account any of the work on planning, discourse, etc. I attribute this fact to the as yet illunderstood process of translation and, thus, the difficulty in defining a basis from which to generate. MT is still struggling with word, phrase and sentence translation (at best) and has therefore perhaps considered it premature to look at discourse problems. The one well-clef'reed problem, and in many ways the most concrete, is the lexicon. Although lexical descriptions imply everything else, there is a feeling that word descriptions and their mapping to other languages can be improved gradually.</Paragraph>
    <Paragraph position="16"> The system is now essentially worked on in Europe and is being extended to other European languages including French, Spanish and Dutch.</Paragraph>
  </Section>
  <Section position="4" start_page="30" end_page="30" type="metho">
    <SectionTitle>
PARTIALLY AUTOMATING THE
TRANSLATION PROCESS
</SectionTitle>
    <Paragraph position="0"> As automation increases and, with it, access to more and more information, the demand for translation increases. Since high-quality machine translation of unrestricted text is no solution to this problem in any foreseeable future, there is a growing trend to look for partial solutions. One option is to build yet another complete MT system such as SYSTRAN, LOGOS, or METAL (the only three viable commercial systems) with full knowledge that the output will be comprehensible, at best. These systems are useful once the lexicon has been developed and tuned for a given corpus; however, given the long development time for building such a system, they will, by definition, be based on out-dated technology. The other solution is to concentrate on those parts of the translation process that can be automated.</Paragraph>
    <Paragraph position="1"> The topic of automating only some parts in view of using a machine to aid in the translation process has been around for a long time (Kay, 1973), often under the name machine or human assisted translation; however, concrete projects addressing a specific aspect are, for the most part, relatively recent. The major aspects currently identified within the space of what can be usefully automated are when and where to use human interaction, identifying classes of restrictions on the input language (lexical, syntactic and semantic), separating the task into sub-tasks (monolingual vs. bilingual and further breaking these down along traditional linguistic lines and according to document preparation criteria). 9 Identifying those aspects which can be automated and working towards a solution has found interest not only from a practical point of view, but also as a theoretical exercise.</Paragraph>
    <Paragraph position="2"> Ambiguity, for example, is an important problem for any language description task. For translation work this problem is compounded in that it may arise not only during parsing, 10 but also during translation and in generation. In generation, the problem is well-known in the CL literature as the problem of natural language paraphrases arising from a given 'meaning' representation; in the context of translation, a system that produces paraphrases is not very useful without some refinement on how the paraphrases differ. Though no formalization of this exists, there is general agreement that paraphrases in a target language often represent different translations w.r.t, a given source text. Another problem for generation in MT and shared with the CL community is that of lexical choice ff the representation abstracts away from words, Seen from a practical viewpoint, the question of automating the process of translation must also take into account the actual working conditions. For example, many translation services work from printed sources rather than the electronic version. And in the larger centers, such as the EC translation services, much of the translation is done via dictaphone and typed by secretaries. In both of these cases, there is no place for interaction unless the entire working pattern is first changed.</Paragraph>
    <Paragraph position="3"> 10 In a recent demonstration I attended of a commercial MT system, it became clear to me that one of the reasons why the output of current systems is so bad, is not that the correct parse couldn't be found among the numerous possible ones, nor that it couldn't be translated, but rather that the system could only choose one parse, and often this was the wrong one.</Paragraph>
    <Paragraph position="4"> representing them as concepts or a set of features. This problem will be apparent in the bilingual component in a transfer system and in generation in an interlingua system.</Paragraph>
    <Paragraph position="5"> One method of controlling the analysis is by limiting the input; one very simple and successful example is the Titus system which has basically a fixed number of templates which define the syntax of the input (Ducrot, 1982). This system, which is actually a database for the textile industry that permits natural language input and provides multilingual output, is also representative of another restriction common in most NLP applications, namely, limiting the application domain.</Paragraph>
    <Paragraph position="6"> Limiting the semantic domain is common practice in the MT community. Eurotra~ for example, works on a limited corpus in the domain of telecommunications and the project reported on in Alshawi, et al., 1991, is developing a lexicon for the ear industry. At ISSCO, work is underway on a sublanguage consisting of avalanche bulletins, somewhat similar to the exemplary TAUM-METEO system for weather reports.</Paragraph>
    <Paragraph position="7"> Another means of automating the translation stages, while still controlling the process, is by allowing for interaction during the various phases (Johnson &amp; Whitelock, 1985).</Paragraph>
    <Paragraph position="8"> Interaction may be limited to the monolingual component where questions are only asked about the source text (Whitelock et al., 1986). The interest in developing authoring systems (i.e., systems where an individual writes a text in the source language with a 'guarantee' that the translation will be correct) is an attempt to assure that the analysis of the text is not ambiguous and does have a translation (McGee Wood &amp; Chandler, 1988). Or the interaction may be included in the bilingual component, where the user is asked to choose the correct translation.</Paragraph>
    <Paragraph position="9"> The simplest example of the latter can be found in the bilingual dictionaries offered with some word processors, often referred to as translator's workstations. Such simple tools as on-line access to both monolingual and bilingual dictionaries, editors that support multiple scrolling screens, or hyphenation and spelling checkers for more than one language are by no means a standard in normal office settings. Development of very simple tools and a basic environment may provide a basis for adding more sophisticated components incrementally.</Paragraph>
    <Paragraph position="10"> Work in MT proper vs. work on environments for translators are essentially carried out in two different eornmunities. Only the commercial systems have up to present taken the latter topic seriously; this situation is changing as more of the funding for research moves to the private sector.</Paragraph>
  </Section>
class="xml-element"></Paper>