Translation Memory vs. Example-based MT – What’s the difference?

Translation Memory vs. Example-based MT – What’s the difference? HAROLD SOMERS (UMIST, Manchester) GABRIELA FERNANDEZ DIAZ (Universidad de Sevilla) 1 Introduction Throughout the 1970s and 1980s, Machine Translation (MT) research focused on the development of so-called “second generation” systems, which aimed to translate text by a process of rule-driven linguistic processing, usually in three stages: syntacticsemantic analysis of the source text, bilingual transfer at a more or less abstract level of representation, and target-text generation from syntactic representation. At the same time, from a practical point of view, there was much discussion on how systems built with this architecture could be used to deliver a reasonable standard of translation usable by real users. Most popular were ideas of restricting the input (the sublanguage and controlled language approaches), or involving the user in preand post-editing. Interactive MT, where the user and the computer would cooperate in resolving ambiguities and making choices, was also championed. In the early 1990s, with these ideas fairly well established, and perhaps even growing stale, research in MT was hit by an apparently new paradigm in which in particular the reliance on linguistic rule systems was to be (at least partially) replaced with the use of a corpus of already-translated examples which would serve as models to the MT system on which to base its new translation. This approach came to be known as Example-Based MT (EBMT) which had in fact first been proposed ten years earlier, in 1981. We will look in a little more detail at this development below. At about the same time, a new tool for translators was being mentioned by developers. Like EBMT, it used a corpus of already-translated examples to serve as models for the new translation, but crucially, it was the human users, not the computer itself, who should determine exactly how to use the examples in producing a new translation. This tool is of course now widely known as a Translation Memory System (TMS). 1 There are many commentators who regard EBMT and TMSs as essentially the same thing, and indeed certain developments in both have brought them closer to each other. The present author has long maintained that there is a crucial difference, which will be elaborated in the next sections. However, there are similarities, and in particular proposals to make TMSs better invariably make them look more and more like EBMT systems. This, principally, is the subject of this paper. 1 We will throughout distinguish the system from its principal component, the database of stored translations, i.e. the “memory”. For this reason we will refer to the systems as TMSs, and to the database itself as a TM. 2 Intertwined history of TMSs and EBMT At a risk of repeating points that are familiar to readers of this journal, we wish in this section to describe some key moments in the development of the concept of TM(S)s. The original idea for a TM is usually attributed to Martin Kay’s well-known “Proper Place” paper (1980), although the details are only hinted at obliquely: [T]he translator might start by issuing a command causing the system to display anything in the store that might be relevant to [the text to be translated] .... Before going on, he can examine past and future fragments of text that contain similar material. (Kay, 1980:19) Interestingly, Kay was pessimistic about any of his ideas for what he called a “Translator’s Amanuensis” ever actually being implemented. But Kay’s observations are actually predated by the suggestion by Peter Arthern (1978) 2 that translators can benefit from on-line access to similar, already translated documents, and in a followup article, Arthern’s proposals quite clearly describe what we now call TM(S)s: It must in fact be possible to produce a programme [sic] which would enable the word processor to ‘remember’ whether any part of a new text typed into it had already been translated, and to fetch this part, together with the translation which had already been translated, .... Any new text would be typed into a word processing station, and as it was being typed, the system would check this text against the earlier texts stored in its memory, together with its translation into all the other official languages [of the European Community]. ... One advantage over machine translation proper would be that all the passages so retrieved would be grammatically correct. In effect, we should be operating an electronic ‘cut and stick’ process which would, according to my calculations, save at least 15 per cent of the time which translators now employ in effectively producing translations. (Arthern, 1981:318). Alan Melby (1995:225f) suggests that the idea might have originated with his group at Brigham Young University (BYU) in the 1970s. What is certain is that the idea was incorporated, in a very limited way, from about 1981 in ALPS , one of the first commercially available MT systems, developed by personnel from BYU. This tool was called “Repetitions Processing”, and was limited to finding exact matches modulo alphanumeric strings. 3 The much more inventive name of “translation memory” does not seem to have come into use until much later. 4 The first TMSs that were actually implemented, apart from the largely inflexible ALPS tool, appear to have been Sumita and Tsutsumi’s (1988) ETOC (“ E asy TO C onsult”), and Sadler and Vendelman’s (1990) Bilingual Knowledge Bank, predating work on corpus alignment which, according to Hutchins (1998) was the prerequisite for effective implementations of the TM idea. Also, Kugler et al. (1991) report on work by Keck (1989) based on statistical methods in the context of an ESPRIT research project. It is difficult to pinpoint when TMs entered the consciousness of translation studies researchers and translators in general. Brian Harris, introducing the notion of a “bi-text” in a translators’ magazine, proposes something like a TM without using that 2 Early proposals for a TM, and other aspects of the idea of a Translator’s Workstation are described in Hutchins (1998). 3 Curiously, this most innovative feature is barely mentioned in descriptions of the ALPS system, two exceptions being Sibley (1988:96,100) and Weaver (1988:121f) who mention the facility almost as an afterthought. 4 Extensive enquiries have so far failed to produce a satisfactory identification of the first use of this term. Hutchins (1998:303) suggests that the Trados company were the first to use the term. name (Harris, 1988:9): 5 a database of paired translations, searchable either by individual word, or by “a whole translation unit”, in the latter case the search being allowed to retrieve similar rather than identical units. Cave (1988) responded to that article with an announcement that Logos were marketing just such a tool. In an unsigned 1991 article the magazine Language International reported that “text banks” had now made their appearance: The first of these would seem to have been IBM European Language Service’s Translation Support Facility (TSF), which, …, incorporated a repeated sentence identification facility. (Anon, 1991:5; emphasis added). The article goes on to explain the notion of “fuzzymatch” (sic) in the case where exact matches are not found. The proceedings of Aslib’s indicative annual conference series Translating and the Computer contain no mention at all of TMs until 1992, when three separate articles (Freibott, 1992; Le-Hong et al., 1992; Svanholm, 1992) mention them, in one case (Le-Hong et al.) without feeling the need to explain the term. Brace (1992) reported development of a TM tool by Trados, as well as the ESPRIT project mentioned above, and projects at IBM’s European Language Services (Denmark) and the Official Languages and Translation sector of the Canadian Department of the Secretary of State in Ottawa. The idea for EBMT has a similar chronology, with ideas surfacing in the early 1980s (the paper presented by Makoto Nagao at a 1981 conference was not published until three years later – Nagao, 1984), but the main developments being reported from about 1990 onwards. 6 The essence of EBMT, called “machine translation by exampleguided inference, or machine translation by the analogy principle” by Nagao, is succinctly captured by his much quoted statement: Man does not translate a simple sentence by doing deep linguistic analysis, rather, Man does translation, first, by properly decomposing an input sentence into certain fragmental phrases ..., then by translating these phrases into other language phrases, and finally by properly composing these fragmental translations into one long sentence. The translation of each fragmental phrase will be done by the analogy translation principle with proper examples as its reference. (Nagao, 1984:178f) Nagao correctly identified the three main components of EBMT: matching fragments against a database of real examples, identifying the corresponding translation fragments, and then recombining these to give the target text. Clearly EBMT involves two important and difficult steps beyond the matching task which it shares with TMS. The idea of EBMT really took off in the early 1990s, with an increasing number of papers at conferences reporting on this approach. Pioneers were mainly in Japan, including Sato and Nagao (1990) and Sumita et al. (1990). Mention should also be made of the work of the DLT group in Utrecht, often ignored in discussions of EBMT, but dating from about the same time as (and probably without knowledge of) Nagao’s work. The matching technique suggested by Nagao involves measuring the semantic proximity of the words, using a thesaurus. A similar idea is found in DLT’s “Linguistic Knowledge Bank” of example phrases described in Pappegaaij et al. 5 In the next paragraph, he describes it as providing “a memory-perfect exploitation of the translator’s own previous experience”. 6 A thorough review of the literature on EBMT is attempted in Somers (1999). (1986a,b) and Schubert (1986:137f). Sadler’s (1991) “Bilingual Knowledge Bank” clearly lies within the EBMT paradigm. During this early period, individual researchers in the field used alternative names, perhaps wanting to bring out some key difference that distinguished their own approach: “case-based” (Collins and Cunningham, 1996), “analogy-based” (Nagao, 1984), and “experience-guided” (Zhao and Tsujii, 1999) are all terms that have been used. The first of these recalls approaches to Machine Learning known as “case-based reasoning” (Riesbeck and Schank, 1989), and other related models. 7 Another term found is “memory-based translation” (Sato and Nagao, 1990; Kitano, 1993), the use of which probably did most to suggest affinities between EBMT and TMSs. 3 What EBMT and TMSs could have in common EBMT and TMSs have in common the use of a database of previous translations, the “memory” or “example-base”, and the essential first step, given a piece of text to translate, of finding in the example database the best match(es) for that text. Once the match has been found, the two techniques begin to diverge. However, it would be misleading to assume that all they have in common is the task of matching, or even that the approaches to matching in the two camps are particularly similar. Use of a database implies issues of database design, content, and maintenance. These will be the focus of the next sections. 3.1 How are examples found? In TMSs, the TM database itself can be constructed in one of three ways. The simplest method, though the most time-consuming one, called “interactive translation” by Bowker (2002:108f) is to build a TM from scratch, that is, to store in the memory each sentence as it is translated. A second method, referred to by Bowker (2002:109f) as “post-translation alignment”, and much heralded by manufacturers, is to extract a TM from an already translated text by aligning the source and target texts. This can be a more or less irksome task (cf. Macdonald, 2001), and there is a considerable literature describing various alignment methods involving differing amounts of (linguistic) sophistication (see Manning and Schütze, 1999:466–486 or Wu, 2000a). O’Brien concludes: A translation memory is always more accurate when it has been created by interactive translation as opposed to automatic alignment, but alignment can produce a reasonably accurate translation memory which can be used as a start-up. (O’Brien, 1998:119) Finally, TMs that have already been created can be imported, and the establishment of agreed interchange formats between manufacturers has hugely facilitated this (notably the Translation Memory eXchange (TMX) format developed by LISA (Localisation Industry Standards Association, cf. Melby, 1998, 2000 and Topping, 2000). The size of the TM is an obvious question. The TM literature says little more than “the bigger the better”, subject to processing limitations, though Bowker (2002:108) warns that “size should not come at the expense of organization”, suggesting that separate TMs for different subject fields or clients may help to reduce false hits caused by homonymy. She adds: 7 The relationship between EBMT and Case-Based reasoning is discussed in Somers and Collins (in press). Keep in mind that a larger TM will result in a greater number of matches .... Therefore, while it may seem logical at first glance to build a single large TM ..., this may turn out to be a false economy.... Moreover, there is a greater likelihood of retrieving “noise” (e.g., matches that are not helpful, matches containing homonyms) and the translator may waste a considerable amount of time analyzing, eliminating, or editing these poor matches. (idem.) For Heyn (1998), a “big” TM will have between 100,000 and 1 million units, thanks to recent technology advances. In recent sparsely-coded-matrix based systems, real interactive work on ‘big’ master translation memories is possible. Big translation memories are typically in the order of 100,000 translation units, although memories in the range of 500,000 to 1,000,000 translation units are envisaged by the end of 1997. According to current research estimates, translation memories could be made up to 40% bigger without any increase in constant access times. (Heyn, 1998:128) In the EBMT literature, the sizes of the example-bases reported vary over a huge range, with 0.73m the biggest, and 7 (seven!) the smallest reported (cf. Somers, 1999:120). Obviously, the systems with tiny example-bases are purely experimental, while the more serious systems will have thousands of examples (rather than, say, hundreds). User manuals for TMSs suggest revising the database every so often to clear out useless examples. By “useless” is presumably meant “not used” rather than, for example, “misleading” (cf. Heyn, 1998:131f). It is easy to see how a TMS could incorporate a measure of the former, simply by counting access. To measure the latter, it would also need to “know” what the translator is doing with the proposed match. The “suitability” of examples is addressed in the context of EBMT systems by various researchers. Nomiyama (1992) introduces the notion of “exceptional examples”, an idea further developed by Watanabe (1994). As far as can be seen, these examples are exceptional just in the sense that if used they give the wrong result! Clearly a more systematic notion is needed. It is well known that the same phrase can be translated differently in different circumstances. Ellipsis, anaphora and stylistic variation can contribute to this, in which case different examples may be seen as nonetheless equivalent in some sense. On the other hand, the underlying meaning of a phrase may differ depending on the context. Somers et al. (1990:274) illustrate how the simple phrase OK in a conversation may be translated into Japanese as wakarimashita ‘I understand’, iidesu yo ‘I agree’ or ijō desu ‘let’s change the subject’. There is also an issue of “granularity”: both in TM and EBMT, there is a tradeoff between length and similarity of examples. The longer the example units, the lower the chance of an exact match; but the shorter the units, the greater the probability of “ambiguity” (multiple, conflicting, matches), with a corresponding decrease in the quality of the proposed translation. Nirenburg et al. (1993:48) call this “passage boundary friction and incorrect chunking”. The obvious and intuitive “grain size” for examples, to judge from almost all TMSs and EBMT systems, is the sentence, though evidence from translation studies (Gerloff, 1987; McTait et al., 1999) suggests otherwise: human translators process text in “naturally-occurring syntactic units” and “generally there is very little processing at sentence level” (McTait et al., 1999). According to Bennett, ... there are good reasons for keeping the U[nit of] T[ranslation] (in the sense of translation atom) in MT as small – and hence as manageable – as possible. Adopting a larger UT may be less efficient, and is not guaranteed to improve translation quality. (Bennett, 1994:18) the “translation atom” being the smallest segment that must be translated as a whole (ibid., p.13). Schäler et al. (2003) echo this sentiment, suggesting that “matching segments at sentence level unnecessarily restricts the potential and the usefulness of translation memories” (p. 89), and propose “phrasal matching” as the primary mechanism for TMSs. Simard (2003) studied how real users make use of a bilingual concordancer, a tool which closely resembles a TMS in function, except that users can look up arbitrary sequences of words. He found that most users look up syntactically well-formed “chunks”, and implemented a system based on this principle. In fact, matching segments rather than whole sentences produces far too many “hits”, so the system must also have a way of selecting the most useful from amongst them. An evaluation of Simard’s implementation suggested that it proposed between 15 and 30 times more “reusable material” than a sentence-based system. Both EBMT and TMSs could probably be improved by concentrating on a more flexible view of the unit of matching/translation, and “the exploitation of fragments of text smaller than sentences” (Cranias et al., 1994:100). In TMSs, this idea is partly addressed in that terminology look-up is often seen as an integral part of the tool, although terminology tools are generally implemented in a lexicon-based rather than a memory-based manner. We will return to the issue of “fragments” below. According to Bowker, Many TM systems allow the user to define other units of segmentation in addition to sentences. These units can include sentence fragments or even entire paragraphs. (Bowker, 2002:94) Along the same lines, Esselink states: A segment is a text element, which is considered by the application as the smallest translatable unit, defined by periods, semi-colons, and hard returns. These are usually sentences, but can also be chapter headings or items in a list. ....Translation memory tools usually allow the user to change and customize segmentation rules. (Esselink, 2000:362f) 3.2 How are examples stored? In TMSs, examples are generally stored as plain text, sometimes with formatting information. Systems differ as to how they treat formatting (i.e. fonts, capitalization and so on) even though it is potentially very useful for matching (see below). Austermühl comments that Some translation memories have a built-in interface that works with common word-processors ... the format in which the translated text is stored in the translation memory is identical to that used in the word-processing program. (Austermühl, 2002:138) From this we must infer that segments always keep their original format when stored. A different issue is the way TMS deal with tags. Nowadays, translation tools are being used extensively in the software localization industry. For this reason, the new generation of TMSs such as Trados, Transit and Déjà Vu contain a wide range of filters to convert files from one format to another. At the same time, TMSs are designed to handle a wide variety of formats such as HTML, SGML and XML. Esselink confirmed the following in the year 2000: Most translation memory tools have standard filters for HTML files. HTML files usually contain very repetitive text, so it is worthwhile using translation memory, because of the substantial time and cost savings. Furthermore, translating updates of web sites is much easier and quicker if a translation memory of the previous version exists. Examples of translation memory tools that support the HTML format are Trados Translator’s Workbench, IBM TranslationManager, STAR Transit, SDLX, and Atril Déjà Vu. (Esselink, 2000:218) In 2003 new versions of TMSs have spread throughout the market. These new versions are well equipped to deal with every kind of application for the design of web pages, presentations, graphics, etc. Thus, Transit XV includes a number of filters that make it possible to translate files generated with programs such as Excel, PowerPoint, QuarkXPress, PageMaker, FrontMaker and AutoCAD, among others. In the same way, the latest solutions presented in the market by Trados and Déjà Vu, namely, Trados 6.5 and Déjà Vu X, incorporate filters which allow the user to import and export files with any kind of format. Increasingly, TMS developers are recognising the value of incorporating “markup” into their systems, not just formatting but also linguistic annotations such as syntactic part-of-speech (POS) tags. In this respect, Planas (1999:8), states that the Xerox XMS Memory Manager, is a “linguistically based tool”, and consequently is capable of retrieving better matches than character-based systems. We read “this shows the crucial importance of using linguistic data for enabling more precise retrieval of the closest sentence in the database”. This system is currently still in the experimental stage however. Planas and Furuse (1999) proposed a much more elaborate scheme in which examples are represented in a multi-level lattice, combining typographic, orthographical, lexical, syntactic and other information. A major drawback of the most successful TMSs available in the market has to do with the lack of incorporating linguistic knowledge in their products. Obviously, storage and matching methods are intricately related: we will return to the latter in the next section. In EBMT systems, a wide range of formats have been proposed for storing the examples. Given its origins as a variant of rule-based MT, early EBMT systems supposed that examples would be stored as aligned tree structures such as the one illustrated in Figure 1, from Watanabe (1992). Figure 1. Representation schema for Kanojo wa kami ga nagai (lit. ‘she TOPIC hair SUBJ is-long’) ↔ She has long hair. Unfortunately, such a representation involves serious overheads in storage space, analysis at run time, and verification of structures, a criticism that also applies to Planas and Furuse’s proposal, as shown in Figure 2. The resulting reliance on parsing or other knowledge-rich processes is acknowledged as a disadvantage. kanojo nagai kami wa ga have she hair long subj obj mod Because these rich representations are widespread in early EBMT proposals, they are often thought of as being a necessary feature of EBMT, though this is quite incorrect. Later EBMT proposals involve much less ambitious representation schemas, in particular lightly annotated text in which words are accompanied by POS tags and/or the results of “stemming” (i.e. morphological analysis to identify root or stem, and partially interpret endings). Planas (1999:8) illustrates the idea by considering sentence (1a) compared to each of (1b-d): although (1c) differs by only one character, humans instinctively find (1b) a better match. (1) a. The white horse is nice. b. The white horses are nice. c. The white house is nice. d. The white houses are nice. Figure 2. Structured representation suggested for ‘Click a color and press ENTER.’ (from Planas and Furuse, 1999, p. 333). An important recent trend in EBMT is to store similar examples in a unified “generalized” manner. For instance, the two examples in (2a,b) could be generalized as (2c), and stored as such, with obvious repercussions for matching (see next section). (2) a. John Miller flew to Frankfurt on December 3rd. b. Dr Howard Johnson flew to Ithaca on 7 April 1997. c. <person-m> flew to <city> on <date> . An early proposal along these lines is found in Furuse and Iida’s (1992) distinction between “literal examples” and “pattern examples”, the latter containing variables in place of words, with the variables characterized by a (list of) typical filler(s), as in (3). 8 8 They also have a third type, called “grammar examples” which consist almost entirely of variables, rather like the rules in conventional MT. (3) a. X o onegai shimasu → may I speak to the X′ (X=jimukyoku ‘office’, …) b. X o onegai shimasu → please give me the X′ (X=bangō ‘number’, …) The idea is quite widespread in the EBMT literature, including Kaji et al.’s (1992) “pseudo-sentences”, Langé et al.’s (1997) “skeleton sentences” and a number of others. 9 The examples are usually generalized by merging similar cases, though authors differ as to whether this can be done (semi-)automatically, or manually. Certainly, if the idea was applied to TMSs, TMs could be much reduced in size, though access would presumably have to be more sophisticated. 3.3 Matching techniques Accessing the TM or example-base involves “matching” the given translation unit against the cases already stored. Early implementations of TMSs could handle only exact matches, although alphanumeric “replaceables” as in (4) were allowed. (4) a. This is shown as A in the diagram. b. This is shown as B in the diagram Bowker introduces the distinction between an “exact” match and a “full” match. An exact match is 100 percent identical to the segment that the translator is currently translating, both linguistically and in terms of formatting. ... This means that the two strings must be identical in every way, including spelling, punctuation, inflection, numbers, and even formatting (e.g., italics, bold). (Bowker, 2002:96f) Thus (5b) would not be retrieved as an exact match for (5a) because of the difference in formatting. (5) a. Click on OK b. Click on OK. A “full match” on the other hand, … occurs when a new source segment differs from a stored TM unit only in terms of so-called variable elements, which are sometimes referred to as “placeables” or “named entities”. Variable elements include numbers, dates, times, currencies, measurements, and sometimes proper names. (ibid., p. 98) Bowker’s “placeables” have been termed “transwords” by Gaussier et al. (1992), while Macklovitch and Russell (2000) call them “non-translatables”. The notion of “named entities” is found in Information Retrieval. As the latter authors point out, they are treated in translation in a rather transparent manner, either not translated at all, or subject to specific conventions, and in any case, independent of context. For a TMS their impact is twofold. On a simple level, we want to have matchers that “ignore” them, so that (4a,b) above are effectively “exact matches”. Additionally, in a more sophisticated TMS (and in EBMT), they are important building blocks for suggesting automatically a likely translation of the given text. Consider (6a) as a text to be translated, which matches with (6b), with the differences highlighted, and its associated translation (6c). In (6b) there are two transwords the “translation” of which can be readily identified in (6c), but you have to know the target language to know which word to change in (6c) to accommodate the lexical difference large vs. small. 9 Nomiyama (1992), Almuallim et al. (1994), Akiba et al. (1995), Collins and Cunningham (1995), Jain et al. (1995), Matsumoto and Kitamura (1995), Watanabe and Takeda (1998), Carl (1999) – see Somers (1999:139ff). elements match up: this can be done by looking at further examples which isolate the words in question. Of course much of this work could be simplified with the help of an on-line dictionary, assuming we had one. But what is of interest to researchers is the extent to which it can be automated. To exemplify this, consider the examples in (19), in a language probably unfamiliar to most readers. (19) a. Dia nak pěrgi kě kědai běli roti. She is going to go to the shops to buy bread. b. Dia pěrgi kě pasar nak běli baju. She went to the market to buy a shirt. c. Měreka pěrgi kě kampung nak běli kereta. They went to the village to buy a car. It is not difficult to identify the probable word-pairings and from the examples to construct the correct translations of sentences like those in (20), and we invite the reader to try it as an exercise. 15 In doing so it should be noted how much generic (common-sense) knowledge about how languages work we as humans bring to this task, which may have to be simulated in an otherwise purely automatic system. (20) a. She went to the village to buy bread. b. They are going to the market. 5 TMSs would be better if they were more like EBMT Our purpose in this paper has been to point out how some of the ideas developed in connection with EBMT could be introduced into the development of TMSs. In this final section we attempt to summarize the main proposals. • If they could identify what in the target part of the match has to be changed. If we want to be able to match similar sentences, such as those differing only in the form of their words, we need linguistic analysis. We do not need a deep analysis that would take a long time to process, but just a “stemming” and a “tagging” one that would give a light but crucial analysis (Planas 1999:9). [T]he most promising strategy for the next generation of TM systems will be to employ various partial parsing or “chunking” techniques. (Macklovitch, 2000) • If they could make suggestions about what in the target part of the match has to be changed to. • If they could construct target texts from matched fragments. • If they could take similar examples and make generalizations about them. TM and EBMT can be seen to lie at opposite ends of a spectrum in memory-based translation. On the one hand, TM requires few linguistic[ ] resources but cannot combine fragments from different T[ranslation] U[nit]s, and on the other hand, EBMT can combine example fragments, but does so by relying on … knowledge-intensive tools. (McTait et al. 1999) 15 The language is Malay. The correct translations are Dia pěrgi kě kampung nak běli roti; Měreka nak pěrgi kě pasar. References Ahrenberg, L. and M. Merkel: 1996. ‘On Translation Corpora and Translation Support Tools: A Project Report’, in K. Aijmer, B. Altenberg, and M. Johansson (eds), Languages in Contrast, Papers from a Symposium on Text-based Cross-linguistic Studies, Lund: Lund University Press, pp. 185–200. Akiba, Y., M. Ishii, H. Almuallim and S. Kaneda: 1995. ‘Learning English Verb Selection Rules from Hand-made Rules and Translation Examples’, in Proceedings of the Sixth International Conference on Theoretical and Methodological Issues in Machine Translation, Leuven, Belgium, pp. 206–220. Al-Adhaileh, M. H. and Tang E. K.: 1998. ‘A Flexible Example-Based Parser Based on the SSTC’, in COLING-ACL ‘98: 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Montreal, Quebec, pp. 687–693. Almuallim, H., Y. Akiba, T. Yamazaki, A. Yokoo and S. Kaneda: 1994. ‘Two Methods for Learning ALT-J/E Translation Rules from Examples and a Semantic Hierarchy’, in COLING 94, The 15th International Conference on Computational Linguistics, Kyoto, Japan, pp. 57–63. Austermühl, F. 2001. Electronic Tools for Translators. Manchester: St. Jerome. Anon: 1991. ‘Bitext makes progress with and without the name’, Language International 3.2:5–6. Arthern, P. J.: 1978. ‘Machine Translation and Computerized Terminology Systems: A Translator’s Viewpoint’, in B.M. Snell (ed.) Translating and the Computer: Proceedings of a Seminar, London, 14th November 1978, Amsterdam (1979): North Holland, pp. 77–108. Arthern, P. J.: 1981. ‘Aids Unlimited: The Scope for Machine Aids in a Large Organization’, Aslib Proceedings, 33:309–319. Bennett, P. 1994. ‘The Translation Unit in Human and Machine’, Babel 40:12–20. Bowker, L. 2002. Computer-Aided Translation Technology. A Practical Introduction. Ottawa: University of Ottawa Press. Brace, C.: 1992. ‘The Many Flavors of Translation Memory’, Language Industry Monitor (May/June 1992); available at http://www.lim.nl/monitor/memory.html. Carl, M.: 1999. ‘Inducing Translation Templates for Example-Based Machine Translation’, in Machine Translation Summit VII, Singapore, pp. 250–258. Carl, M.: 2001. ‘Inducing Translation Grammars from Bracketed Alignments’, in MT Summit VIII Workshop on Example-Based Machine Translation, Santiago de Compostela, Spain, pp. 12–23; repr. in Carl and Way (2003), pp. 339–361. Carl, M. and A. Way (eds): 2003. Example-Based Machine Translation, Dordrecht: Kluwer. Cave, J.R.: 1988. ‘Observations on Bi-Text’, Letter to the Editor, Language Monthly 57:18–19. Chatterjee, N.: 2001. ‘A Statistical Approach for Similarity Measurement Between Sentences for EBMT’, in STRANS–2001 Symposium on Trabnslation Support Systems, Kanpur, India. [no page numbers] Cicekli, I., and H. A. Güvenir: 1996. ‘Learning Translation Rules From A Bilingual Corpus’, in NeMLaP-2: Proceedings of the Second International Conference on New Methods in Language Processing, Ankara, Turkey, pp. 90–97; repr. as ‘Learning Translation Templates from Bilingual Translation Examples’ in Carl and Way (2003), pp. 255–286. Collins, B.: 1998. Example-Based Machine Translation: An Adaptation-Guided Retrieval Approach. PhD thesis, Trinity College, Dublin. Collins, B. and P. Cunningham: 1995. ‘A Methodology for Example Based Machine Translation’, in CSNLP 1995: 4th Conference on the Cognitive Science of Natural Language Processing, Dublin. Collins, B. and P. Cunningham: 1996. ‘Adaptation-Guided Retrieval in EBMT: A Case-Based Approach to Machine Translation’, in I. Smith and B. Faltings (eds), Advances in Case-Based Reasoning: Third European Workshop, EWCBR-96, Berlin: Springer, pp. 91–104. Cranias, L., H. Papageorgiou and S. Piperidis: 1997. ‘Example Retrieval from a Translation Memory’, Natural Language Engineering 3:255–277. Dennett, G.: 1995. Translation memory: Concept, products, impact and prospects. MSc dissertation, School of Electrical, Electronic and Information Engineering, South Bank University, London. Esselink, B. 2000. A Practical Guide to Localization, 2nd ed. Amsterdam: John Benjamins. Frederking, R., S. Nirenburg, D. Farwell, S. Helmreich, E. Hovy, K. Knight, S. Beale, C. Domashnev, D. Attardo, D. Grannes and R. Brown: 1994. ‘Integrating Translations from Multiple Sources within the Pangloss Mark III Machine Translation System’, in Technology Partnerships for Crossing the Language Barrier: Proceedings of the First Conference of the Association for Machine Translation in the Americas, Columbia, Maryland, pp. 73–80. Freibott, G.P.: 1992. ‘Computer Aided Translation in an Integrated Document Production Process: Tools and Applications’, in Translating and the Computer14: Quality Standards and the Implementation of Technology in Translation, London, pp. 15–24. Furuse, O. and H. Iida: 1994. ‘Constituent Boundary Parsing for Example-Based Machine Translation’, in COLING 94, The 15th International Conference on Computational Linguistics, Kyoto, Japan, pp. 105–111. Gaussier, E., J.-M. Langé and F. Meunier: 1992. ‘Towards bilingual terminology’, in Proceedings of the ALLC/ACH Conference, Oxford, pp. 121–124. Gerloff, P.: 1987. ‘Identifying the Unit of Analysis in Translation’, in C. Færch and G. Kasper (eds), Introspection in Second Language Research, Clevedon: Multilingual Matters, pp. 135–158. Güvenir, H. A. and I. Cicekli: 1998. ‘Learning Translation Templates from Examples’, Information Systems 23:353–363. Harris, B.: 1988. ‘Bi-text, a new concept in translation theory’, Language Monthly 54:8–10. Heyn, M. 1998: ‘Translation Memories – Insights & Prospects’, in L. Bowker, M. Cronin, D. Kenny and J. Pearson (eds) Unity in Diversity? Current Trends in Translation Studies, Manchester: St Jerome, pp. 123–136. Hutchins, J. 1998: ‘The Origins of the Translator’s Workstation’. Machine Translation, 13:287–307. Jain, R., R. M. K. Sinha and A. Jain: 1995. ‘Role of Examples in Translation’, in 1995 IEEE International Conference on Systems, Man and Cybernetics, Vancouver, BC, pp. 1615–1620. Kaji, H., Y. Kida and Y. Morimoto: 1992. ‘Learning Translation Templates from Bilingual Text’, in Proceedings of the fifteenth [sic] International Conference on Computational Linguistics, Actes du quinzième colloque international en linguistique informatique, COLING-92, Nantes, France, pp. 672–678. Kay, M.: 1980. The Proper Place of Men and Machines in Language Translation. Research Report CSL-80-11, Xerox PARC, Palo Alto, Calif. Reprinted in Machine Translation 12:3–23 (1997). Keck, B.: 1989. ‘Theoretical Study of a Statistical Approach to Translation’, TWB Technical Report, Fraunhofer Institute IAO, Stuttgart, Germany. Kitano, H.: 1993. ‘A Comprehensive and Practical Model of Memory-Based Machine Translation’, in Proceedings of the International Joint Conference on Artificial Intelligence, Chambéry, France, pp. 1276–1282. Kugler, M., G. Heyer, R. Kese, B. von Kleist-Retzow and G. Winkelmann: 1991. ‘The Translator’s Workbench: An Environment for Multi-Lingual Text Processing and Translation’, in Machine Translation Summit III Proceedings, Washington, DC, pp. 81–83. Langé, J.-M., É. Gaussier and B. Daille: 1997. ‘Bricks and Skeletons: Some Ideas for the Near Future of MAHT’, Machine Translation 12:39–51. Le-Hong, K., M. Höge and A. Hohman: 1992. ‘User’s Point of View of the Translator’s Workbench’, in Translating and the Computer14: Quality Standards and the Implementation of Technology in Translation, London, pp. 25–32. Macdonald, K.: 2001. ‘Improving Automatic Alignment for Translation Memory Creation’, in Translating and the Computer 23: Proceedings from the Aslib Conference held on 29 & 30 November 2001, London, [pages not numbered]. Macklovitch, E.: 2000. ‘Two Types of Translation Memory’, in Translating and the Computer 22: Proceedings from the Aslib Conference 16 & 17 November 2000, London, [pages not numbered]. Macklovitch, E. and G. Russell: 2000. ‘What’s Been Forgotten in Translation Memory’, in J. S. White (ed.) Envisioning Machine Translation in the Information Future: 4th Conference of the Association for Machine Translation in the Americas, AMTA 2000, Berlin: Springer, pp. 137– 146. Manning, C. D. and H. Schütze: 1999. Foundations of Statistical Natural Language Processing. Cambridge, Massachusetts: The MIT Press. Maruyama, H. and H. Watanabe: 1992. ‘Tree Cover Search Algorithm for Example-Based Translation’, in Fourth International Conference on Theoretical and Methodological Issues in Machine Translation. Empiricist vs. Rationalist Methods in MT. TMI-92, Montréal, Québec, pp. 173–184. Matsumoto, Y., H. Ishimoto and T. Utsuro: 1993. ‘Structural Matching of Parallel Texts’, in 31st Annual Meeting of the Association for Computational Linguistics, Columbus, Ohio, pp. 23–30. Matsumoto, Y. and M. Kitamura: 1995. ‘Acquisition of Translation Rules from Parallel Corpora’, in R. Mitkov and N. Nicolov (eds) Recent Advances in Natural Language Processing: Selected Papers from RANLP’95, Amsterdam: John Benjamins, pp. 405–416. McTait, K.: 2001. ‘Linguistic Knowledge and Complexity in an EBMT System Based on Translation Patterns’, in MT Summit VIII Workshop on Example-Based Machine Translation, Santiago de Compostela, Spain, pp. 23–34; to be repr. Carl and Way (2003). McTait, K., M. Olohan and A. Trujillo: 1999. ‘A Building Blocks Approach to Translation Memory’, in Translating and the Computer 21: Proceedings from the Aslib Conference, London, [pages not numbered]. Melby, A. K.: 1995. The Possibility of Language: A Discussion of the Nature of Language. Amsterdam: John Benjamins. Melby, A. K.: 1998. ‘Behind the Scenes: Data-exchange standards are unsung heroes revolutionizing the language industries’, Language International 10.6:30–31, 43. Melby, A. K.: 2000. ‘Sharing of translation memory databases derived from aligned parallel text’, in Véronis (2000), pp. 347–368. Merkel, M., B. Nilsson and L. Ahrenberg: 1994. ‘A phrase-retrieval system based on recurrence’, in Second Annual Workshop on Very Large Corpora (WVLC2), Kyoto, Japan, pp. 99–108. Nagao, M.: 1984. ‘A Framework of a Mechanical Translation between Japanese and English by Analogy Principle’, in A. Elithorn and R. Banerji (eds) Artificial and Human Intelligence, Amsterdam: North-Holland, pp. 173–180. Nirenburg, S., C. Domashnev and D. J. Grannes: 1993. ‘Two Approaches to Matching in ExampleBased Machine Translation’, in Proceedings of the Fifth International Conference on Theoretical and Methodological Issues in Machine Translation TMI ‘93: MT in the Next Generation, Kyoto, Japan, pp. 47–57. Nomiyama, H.: 1992. ‘Machine Translation by Case Generalization’, in Proceedings of the fifteenth [sic] International Conference on Computational Linguistics, COLING-92, Nantes, France, pp. 714–720. O’Brien, S. 1998: ‘Practical Experience of Computer-Aided Translation Tools in the Software Localization Industry’, in L. Bowker, M. Cronin, D. Kenny and J. Pearson (eds) Unity in Diversity? Current Trends in Translation Studies, Manchester: St Jerome, pp. 115–122. Pappegaaij, B. C., V. Sadler and A. P. M. Witkam (eds): 1986a. Word Expert Semantics: An Interlingual Knowledge-Based Approach. Dordrecht: Reidel. Pappegaaij, B. C., V. Sadler and A. P. M. Witkam: 1986b. ‘Experiments with an MT-Directed Lexical Knowledge Bank’, in 11th International Conference on Computational Linguistics: Proceedings of Coling ’86, Bonn, West Germany, pp. 432–434. Planas, E.: 1999. ‘Towards Second-Generation Translation Memories’, LISA Newsletter 8.2:7–11. Planas, E. and O. Furuse: 1999. ‘Formalizing Translation Memories’, in Machine Translation Summit VII, Singapore, pp. 331–339; repr. in Carl and Way (2003), pp. 157–188. Rapp, R. 2002. ‘A Part-of-Speech-Based Search Algorithm for Translation Memories’ in LREC 2002, Third International Conference on Language Resources and Evaluation, Las Palmas de Gran Canaria, Spain, pp. 466–472. Riesbeck, C. and R. Schank: 1989. Inside Case-Based Reasoning. Hillsdale, NJ: Lawrence Erlbaum. Sadler, V.: 1991. ‘The Textual Knowledge Bank: Design, Construction, Applications’, in International Workshop on Fundamental Research for the Future Generation of Natural Language Processing (FGNLP), Kyoto, Japan, pp. 17–32. Sadler, V. and R. Vendelmans: 1990. ‘Pilot Implementation of a Bilingual Knowledge Bank’, in COLING-90, Papers Presented to the 13th International Conference on Computational Linguistics, Helsinki, Finland, Vol. 3, pp. 449–451. Sato, S. and M. Nagao: 1990. ‘Toward Memory-Based Translation’, in COLING-90, Papers Presented to the 13th International Conference on Computational Linguistics, Helsinki, Finland, Vol. 3, pp. 247–252. Schäler, R., A. Way and M. Carl: 2003. ‘EBMT in a controlled environment’, in Carl and Way (2003), pp. 83–114. Schubert, K.: 1986. ‘Linguistic and Extra-Linguistic Knowledge: A Catalogue of Language-related Rules and their Computational Application in Machine Translation’, Computers and Translation 1:125–152. Sibley, J.: 1988. ‘Le système ALPS’, in Actes du Séminaire internationale “La traduction assistée par ordinateur”, Perspectives technologiques, industrielles et économiques envisageables à l’horizon 1990”, Paris, pp. 95–102. Simard, M.: 2003. Mémoires de traduction sous-phrastiques. Thèse de doctorat, Université de Montréal. Simard, M. and P. Langlais: 2001. ‘Sub-sentential Exploitation of Translation memories’, in MT Summit VIII: Machine Translation in the Information Age, Santiago de Compostela, Spain, pp. 335–339. Somers, H.: 1999. ‘Review Article: Example-based Machine Translation’, Machine Translation 14:113–157; revised version in Carl and Way (2003), pp. 3–57. Somers, H., I. McLean and D. Jones: 1994. ‘Experiments in Multilingual Example-Based Generation’, in CSNLP 1994: 3rd Conference on the Cognitive Science of Natural Language Processing, Dublin, Ireland, [pages not numbered]. Somers, H., J. Tsujii and D. Jones: 1990. ‘Machine Translation without a Source Text’, in COLING-90, Papers Presented to the 13th International Conference on Computational Linguistics, Helsinki, Finland, Vol. 3, pp. 271–276. Sumita, E. and H. Iida: 1995. ‘Heterogeneous Computing for Example-Based Translation of Spoken Language’, in Proceedings of the Sixth International Conference on Theoretical and Methodological Issues in Machine Translation, Leuven, Belgium, pp. 273–286. Sumita, E., H. Iida and H. Kohyama: 1990. ‘Translating with Examples: A New Approach to Machine Translation’, in The Third International Conference on Theoretical and Methodological Issues in Machine Translation of Natural Language, Austin, Texas, pp. 203–212. Sumita, E. and Y. Tsutsumi: 1988. ‘A Translation Aid System Using Flexible Text Retrieval Based on Syntax-Matching’, in Second International Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages, Proceedings Supplement, Pittsburgh, Pennsylvania, [pages not numbered]. Svanholm, F.: 1992. ‘“The Happy Triad” – The Human, the MAT, and the MT’, in Translating and the Computer14: Quality Standards and the Implementation of Technology in Translation, London, pp. 15–24. Topping, S.: 2000. ‘Sharing Translation Database Information: Considerations for developing an ethical and viable exchange of data’, MultiLingual Computing & Technology 11.5:59–61. Veale, T. and A. Way: 1997. ‘Gaijin: A Bootstrapping Approach to Example-Based Machine Translation’, International Conference, Recent Advances in Natural Language Processing, Tzigov Chark, Bulgaria, pp. 239–244. Véronis, J. (ed.): 2000. Parallel text processing: Alignment and Use of Translation Corpora. Dordrecht: Kluwer. Watanabe, H.: 1994. ‘A Method for Distinguishing Exceptional and General Examples in ExampleBased Transfer Systems’, in COLING 94, The 15th International Conference on Computational Linguistics, Kyoto, Japan, pp. 292–301. Watanabe, H. and K. Takeda: 1998. ‘A Pattern-Based Machine Translation System Extended by Example-based Processing’, in COLING-ACL ’98: 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Montreal, Quebec, pp. 1369–1373. Way, A.: 2001. ‘Translating with Examples’, in MT Summit VIII Workshop on Example-Based Machine Translation, Santiago de Compostela, Spain, pp. 66–80; repr. in Carl and Way (2003), pp. 443–472. Weaver, A.: 1988. ‘Two Aspects of Interactive Machine Translation’, in M. Vasconcellos (ed.) Technology as Translation Strategy, Binghamton, NY: State University of New York at Binghamton (SUNY), pp. 116–123. Wu, D.: 2000a. ‘Alignment’, in R. Dale, H. Moisl and H. Somers (eds) Handbook of Natural Language Processing, New York: Marcel Dekker Inc., pp. 415–458. Wu, D.: 2000b. ‘Bracketing and aligning words and constituents in parallel text using Stochastic Inversion Transduction Grammars’, in Véronis (2000), pp. 139–167. Zhao, G. and J. Tsujii: 1999. ‘Transfer in Experience-Guided Machine Translation’, in Machine Translation Summit VII, Singapore, pp. 501–508.