Parallel Corpora in translator education

PARALLEL CORPORA IN TRANSLATOR EDUCATION GUADALUPE RUIZ YEPES UNIVERSITÄT HILDESHEIM ABSTRACT Mainly four kinds of corpora have been used in the field of translator education: comparable corpora, parallel corpora, parallel learner corpora and ‘do it yourself’ or ‘disposable’ corpora. This article focuses on the use of parallel corpora in translator education. As a starting point different approaches to translator education with corpora will be described in order to figure out which one has made use of a methodological background in translation didactics to carry out their research. Secondly different parallel corpora will be presented and the strengths and limitations of different tools, such as bilingual concordancers and translation memories, will be analysed. To round off this article different translation alternatives to the same source text will be explored using EMEA in the context of translators’ training. KEYWORDS: corpus linguistics, translator education CONTENTS: 1. Introduction 2. Parallel corpora 2.1. Alignment 2.2. Parallel corpora available in the web 2.3. Tools 2.3.1. Bilingual/Multilingual Concorancers 2.3.2. Translation Memories 2.3.3. TMs and BCs/MCs in translator training/education 2.4. Learning to observe and translate patterns with the help of parallel corpus: EMEA (European Medicines Agency documents) 3. Conclusions 4. Bibliography UHGLWQ~PHUR Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU 1. Introduction The use of corpora has become very popular during the last decades in translation education. The most used so far are comparable, parallel, parallel learner and ‘do it yourself’or ‘disposable’ corpora. Comparable corpora1 are the most used in the field of translator education simply because the accessibility to monolingual texts in the web makes their compilation less time-consuming than the compilation of parallel corpora2. Many Translation lecturers encourage their students to use comparable corpora as they would use dictionaries; to look up terminology or lexicogrammatical phenomena such as collocations and semantic prosodies. Kübler (cf. 2003:35), for instance, explains how students can use comparable corpora in order to figure out which adjectives modify certain verbs in the field of cameras. Cosme, on her part, applies comparable corpora for English and French to make students analyse and compare hypotaxis and parataxis in these two languages (cf. 2006: 97). As for parallel corpora, due to the difficulty of finding texts with their translations in the web, most lecturers use existing corpora in their translation courses instead of trying to compile their own. Marco and Van Lawick (2009), for instance, use COVALT (Corpus Valenciano de Literatura Traducida) to design translation exercises (cf. 2009: 13-16). Kübler’s students had access to a parallel corpus composed by Linux HowTo’s original texts with their translations into French. This corpus allowed them to figure out that some English terms have to be translated into French in a different way depending on the text genre. In a French specialized journal, for instance, the word firewall is supposed to be left in English while in an operating manual intended for laymen the term would be translated into French as garde-barrière or coupe-feu (cf. 2003:39). There are many reasons that evidence the usefulness of parallel learner corpora for Translation lecturers. By comparing translations of the same original text by different students in the same course the lecturer can detect what translation problems caused them more difficulties. By comparing the translations of different original texts by a single student the lecturer can observe the learning progression of that particular student. Since learner corpora have been used in the context of second language acquisition with great success (cf. Granger 1993, 1998), Bowker and Bennison decided to compile a corpus of digitalized translations by students and called it Student Translation Archive (STA). They developed for that special corpus a query language; Student Translation Tracking System (STTS) (cf. Bowker and Bennison 2003:104). Their main motive being to build a corpus that is able to “provide a means of identifying areas of difficulty that could then be integrated into the curriculum and discussed in class” (cf. 2003:103). Following this line, the MeLLANGE (Multilingual eLearning in LANGuage Engineering) project was launched in order to create a multilingual annotated aligned corpus of translations produced by translation students and professionals. This Learner Translator Corpus (LTC) was annotated with part-ofspeech tags and lemma information, as well as error annotations to identify and classify types of errors. The meta-data concerning the translator is also recorded because this 1 “What links the collections of texts in comparable corpora is that they have been put together according to the same type of criteria (texts of a certain size, on a set topic, from a given period, etc.)“(Kenning 2010: 487). 2 “The prototypical parallel corpus consists of a set of texts in language A and their translations in language B (e.g. the works of Dickens and their translations into French).“ (Kenning 2010: 487). redit, número 7, 2011, - 66 information may help lecturers to associate types of errors with text production circumstances. A similar translation collection mechanism to the one designed by Bowker and Bennison (cf. 2003: 103-118) was applied in order to collect the meta-data. Last but not least, disposable corpora ought to be mentioned. They are also known as DIY (do it yourself) corpora and are usually compiled for the fulfillment of one translation assignment (cf. Varantola 2003:55). Disposable corpora are usually comparable corpora with a simple structure composed of texts downloaded from the internet and designed to solve the translation problems of one particular translation assignment. According to Varantola, the know-how about compiling a disposable corpus is part of the translation competence and should be regarded as fundamental in the training of translators: As an umbrella term, we could use corpus management in translation and by this refer to the knowledge and skills needed in the compilation and use of corpus information for individual translation assignments. (Varantola 2003:57). In this sense the group WaCky3 has been carrying out significant research. Their main purpose is to deal with the topic Web as Corpus (WaC) and to find out how the web can help us compiling a corpus. Baroni and Bernardini (cf. 2006:10) focus on four different ways of using the web as corpus: the Web as corpus surrogate, the Web as a corpus shop, the Web as corpus proper and the Web as mega-corpus/mini-web. The studies mentioned so far are extremely practice-oriented. They describe situations and exercises carried out in class without referring to any theories or methodological backgrounds although there are at least 7 different major methodological approaches to translator education (cf. Kelly 2005: 11-18): - Objectives-centred approach: Delisle - Profession-based and learner-centred approach: Nord - The socioconstructive approach: Kiraly - The psycholinguistic approach: Hönig and Kussmaul - Process-centred approach: Gile - Task-based approach: Hurtado, Gónzalez Davies - The situational approach: Gouadec The PACTE group is an exception in this respect as they carry out research applying corpora in the teaching of translation within the task-oriented approach based on Delisle’s postulates. Delisle was probably the first translation scholar to establish teaching objectives for translator education (cf. Kelly 2005: 11). In his book L’analyse du discours comme méthode de traduction (1980) he suggests 23 teaching objectives which can be general or specific. Hurtado Albir (cf. 1999) and González Davies (cf. 2004) follow this research stream and develop translation curricula that are based on teaching objectives. Marco Borrillo and Van Lawik (cf. 2009: 16) design, within the PACTE group, corpus based and corpus driven multiple choice and cloze test exercises for translator education using mainly comparable and parallel corpora. In the next chapters of this article the use of parallel corpora in the context of translator education will be discussed. Different corpora such as the OPUS corpora collection and BancTrad will be presented. The strengths and limitations of different tools, such as bilingual concordancers and translation memories, will be analysed and the usefulness of parallel corpora in translators’ education will be highlighted. 3 Web-as-Corpus kool yniative. Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 67 2. Parallel corpora 2.1. Alignment While the comparison between corpora of original texts in different languages has been a very important domain of contrastive linguistics, the comparison of translated texts in different languages and their originals has been favoured in the field of corpus-based translations studies. Nevertheless, comparable corpora have been the most used in translator education, which has to do with the fact that parallel corpora are much more expensive to build. They must be aligned in order to allow observing the existence of equivalent text segments in the languages being analysed. Although there are three levels of alignment: paragraph, sentence and word, the sentence unit is the most used as the basic alignment segment. That does not mean that each sentence in language A is aligned with one sentence in language B throughout the texts, since two sentences in language A could have been translated into just one sentence in language B or vice versa. If the paragraph is taken as the basic alignment unit the results of a search are more difficult to analyse because the searched words and their translations will be embedded within a much larger amount of text. The usual alignment procedure involves using alignment software to align texts automatically and inspecting the output manually in order to correct misalignments. The wide range of alignment software includes tools integrated in translation memory systems such as SDL Trados WinAlign and in parallel concordancers such as Multiconcord. However, there are also some ‘independent’ aligners such as Danielsson and Ridings’ (1997) ‘vanilla’ aligner or the Microsoft bilingual sentence aligner (cf. Kenning 2010: 490-491). Table 1 shows two examples of sentence alignment from the EMEA corpus for the search of the German word ‘Blut’ with their equivalents in English and Spanish. de en es 3015064 Daher werden Episoden mit hohen Kalziumwerten im Blut gegenüber der alleinigen Einnahme von Kalziumtabletten zur Regulierung des Phosphatspiegels im Blut reduziert So the number of episodes of high calcium levels in your blood will be reduced as compared to when you are taking calcium tablets alone to regulate the phosphate levels in the blood. Por lo tanto, el número de episodios de niveles de calcio altos en sangre se verá reducido en comparación con el número de episodios cuando se toman comprimidos de calcio por sí solos, utilizados para regular los niveles de fosfato en sangre. 9075905 Die häufigsten Nebenwirkungen (beobachtet bei 1 bis 10 The most common sideeffects with Sebivo (seen in between 1 and 10 Los efectos secundarios más frecuentes de Sebivo (observados entre 1 y 10 Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 68 von 100 Patienten) sind Schwindel, Kopfschmerzen, Husten, erhöhte Konzentrationen einziger Enzyme im Blut (Amylase, Lipase, Kreatin-Phosphokinase, Alanin-Aminotransferase), Durchfall, Übelkeit (Unwohlsein), Bauchschmerzen, Hautausschlag und Erschöpfung (Müdigkeit). patients in 100) are dizziness, headache, cough, increased levels of some enzymes in the blood (amylase, lipase, creatine phosphokinase, alanine aminotransferase), diarrhoea, nausea (feeling sick), abdominal (tummy) pain, skin rash and fatigue (tiredness). pacientes de cada 100) son mareo, tos, incremento de los niveles séricos (en sangre) de algunas enzimas (amilasa, lipasa, creatinina, fosfoquinasa, alanina aminotransferasa ALT), diarrea, náuseas, dolor abdominal, prurito cutáneo y fatiga (cansancio). Table 1: An example of sentence alignment from the EMEA corpus. 2.2. Parallel corpora available in the web Existing parallel corpora tend to be domain-specific (mainly parliamentary proceedings and technical manuals), considerably small and are scarcely annotated. It is for these reasons that the following corpora deserve special consideration (cf. Xiao 2009): •JRC-ACQUIS (Multilingual Parallel Corpus) comprises texts in 22 languages about the European legislation (over one billion tokens in total). 231 language combinations are aligned. These corpora are accessible free of charges for researchers over the Acquis website. •JOC (Official Journal of European Community) comprises annotated texts (postags) in five languages (English, German, Italian, Spanish and French). 4 language combinations are aligned (English-German, English-Italian, EnglishSpanish and English-French). In this case only some corpora are accessible free of charge. •MLCC (Multilingual Corpora for Cooperation) comprises a comparable (Comparable Polylingual Document collection) and a parallel corpus (Multilingual Parallel Corpus). The parallel corpus is composed of texts in nine languages from the Official Journal of European Community. This corpus is accessible through ELRA. •OPUS parallel corpus is composed of European and Assian translated texts from the internet. It comprises hitherto 352 millions of words, but it is growing incessantly because it is being fed constantly with new texts. OPUS is accessible free of charges and allows access to corpora such as OO (the OpenOffice), EUconst (The European constitution) EMEA (European Medicines Agency documents) and EUROPARL (European Parliament Proceedings). •CRATER Corpus (The Corpus Resources and Terminology Extraction project) comprises annotated and aligned texts about international telecommunications in three languages: Spanish, French and English (one million tokens per language). •BancTrad (BT) is a web interface for access to annotated monolingual and parallel corpora. The languages included in the parallel corpora are Catalan, Spanish, English, German and French. The language combinations of Catalan and Spanish with the other three languages are aligned. Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 69 Other examples are COMPARA (Portuguese, English), INTERSECT (English, French and German), etc. To mention them all would go beyond the scope of this paper. The Linguistic Data Consortium, hosted by the University of Pennsylvania, offers a catalogue of ready-made parallel corpora available for researchers liable to pay charges (see www.ldc.upenn.edu). The efficiency of automated retrieval from a corpus depends greatly on the levels of annotations included and, therefore, the types of queries allowed. The possible linguistic levels of annotation in corpora so far are the orthographic, the phonetic, the prosodic, the semantic, the morphologic, the syntactic and the pragmatic levels. However, ideally a corpus should include at least these four levels of annotation: words (tokenization), lemmata (lemmatization), morphologic (Part-of-speech tags) and syntantic function (parsing). Corpora interfaces also influence information retrieval from corpora. One of the main hazards impeding the quick propagation of work with corpora among translation professionals is the fact that too much time has to be spent in order to become familiar with the different interfaces and query languages (cf. Colominas and Badia 2008: 76). Translation training institutions are aware of this problem and are implementing their systems by creating uniform interface platforms from which researchers, lecturers and students can access more than one corpus applying just one query language. This is the case for several parallel corpora; for instance, the OPUS website puts CQP (Corpus Query Processor) at the disposal of its visitors for the retrieval from its subcorpora. 2.3. Tools Acccording to Bowker and Barlow there are two main types of tools “that can be used to search for and retrieve information from a bilingual parallel corpus: a bilingual concordancer (BC) and a translation memory (TM)” (2008: 2). While TMs have enjoyed always widespread recognition within the translation industry, BCs don’t seem to be found useful by professional translators and are used rather in academic circles. Bernardini and Castagnoli (cf. 2008: 48) admit that corpus use is not only timeconsuming but also cognitively demanding and professional translators always work under time constraints and are glad to apply user-friendlier resources. Nevertheless, it is indisputable that both tools have strengths and limitations, so that in certain situations the use of a BC could be more enlightening for a translator trying to figure out, for instance, possible semantic prosodies of words. It is not the aim of this article to compare BCs with TMs since Bowker and Barlow (2008: 1-22) have done so already with ParaConc and SDL Trados. The main goal of this chapter is, firstly, to present briefly different BCs and TMs in order to gain an overview of available types of tools to retrieve information from parallel corpora and, secondly, to describe to what extent have they been applied in the context of translator education. 2.3.1. Bilingual/Multilingual Concordancers Most concordancers are not tied to particular languages. They are hence multilingual rather than bilingual, since they are designed to work with a pair of languages (corpora composed by texts in a source language with its translation into one language), or with Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 70 various languages, (corpora composed by texts in a source language with its translations into several languages). They should be called, therefore, Multilingual Concordancers (MCs). MCs allow translators to search through parallel corpora “to find information that might help them to complete a new translation” (ibd. 2008: 3). But as already mentioned, professional translation is not the only field of application of MCs; they can also be applied in the fields of contrastive analyses, language learning and translator training/education. These are the most well-known MCs on the market: •ParaConc: its alignment takes the sentence as basic unit and is based on the Gale-Church algorithm. ParaConc offers a search feature for the analysis of parallel texts, which allows observing the search results in a KWIC layout in both languages, to sort the concordance lines and to move from one language to another. Other utilities of ParaConc are ‘hotwords’, ‘frequency information’ and ‘advanced search’. •TransSearch: is a web-based service which also takes the sentence as basic alignment unit. It provides access to a translation database containing millions of sentences translated into English, French and Spanish. TranSearch arranges search results into two columns: the first containing sentences of the source texts in which the word or expression occurs, and the second containing the translation of those sentences. It is a search engine that could be compared to Google, but it is applied to source texts with its translations. •Find Bi-Text Advantage: Beetext Find has put on the market three versions of this concordancer: for freelancers (Desktop Find Bi-Text Advantage - DBA), for translators employed by small businesses (Small Business Find Bi-Text Advantage - SBBA) and for translators employed by large enterprises (Enterprise Find Bi-Text Advantage - EBA). SBBA and EBA can be networked to allow several members of a team work with the same concordancer and EBA allows users to see documents in their original layout with their graphics and not just as plain texts. •MultiConcord: is a multilingual concordancer for Windows by Lingua, which combines a parallel concordancer with facilities for the semiautomatic generation of texts. Users are able to add their own pairs of texts to the corpus, using simple mark-up conventions based on SGML. 2.3.2. Translation Memories The reuse of text segments that have been previously translated in the past is the main function of translation memory systems. The underlying principle is to save every translated sentence in a database creating a parallel corpus of aligned translation units and to compare each new sentence that needs to be translated with this database in order to reuse existing translations. There are many different TMs on the market and the recycling of existing translations is their common feature, but we are going to focus now on what differentiates them: •SDL Trados Studio: offers, apart from the usual translation memories features, the AutoSuggest (smart suggestions as you type) and the Automated Translation feature (if the translation memory does not provide a match for a particular phrase, this feature helps to translate it). Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 71 •Déjà Vu X2: includes different features such as the DeepMiner (carries out cross-analyses of the databases you have built to “mine” translations and phrase segments in them), AutoWrite (an interactive predictive translation system even for texts with few or no database segment matches that proposes phrases and sentences mined from your database as you translate), Guaranteed Matches (takes the context of each segment into account to select the most appropriate match), Project Lexicon (offers temporary glossaries) and AutoSearch (shows not just relevant matches but also similar phrases and segments) among other things. •STAR Transit: In contrast to other TM systems, Transit does not use an exclusively sentence based database. Instead, the context contained in a document remains available in the reference material. •WordFast: Has been designed to run on multiple platforms, including Windows, Linux, and Mac. Wordfast is filled with features such as real time quality assurance and the ability to hook into multiple glossaries of any format. •Across: As SDL Trados Studio and Déjà Vu X2, Across offers a feature for the intelligent automatic completion of sentences during translation, based on the source text and the keystrokes entered by the translator. 2.3.3. TMs and BCs/MCs in translator training/education Most TM providers are interested in becoming popular within the educational community, as they are aware of the fact that a high percentage of translation students will be professional translators in a near future and possible customers for them. Providers try to promote their TMs as the ideal product in the learning environment and some of them even list in their websites the Educational Institutions that buy their products. In the last decades it has become evident that most translators need certain technological tools to practise their work. This is especially the case for those working in the technical field and therefore translating operating manuals, software-surfaces, online-helps and so forth. In order to avoid having discrepancies between what students learn at universities and the requirements of the profession they are being trained for, universities are adapting their curricula to the demands of the market. Some institutions dedicated to the training of translators like the University of Mainz/Germersheim, the School for Applied Linguistics and Translatology (Institut für angewandte Linguistik und Translatologie) at the University of Leipzig or the University Pompeu Fabra in Spain followed the LETRAC (Language Engineering for Translator Curricula, LE4-8324) project, which was financed by the EU and created in order to design a common language engineering curricula for Universities. Translation curricula at universities focus therefore more and more on ‘training’ students for their future as professional translators and since TMs are the most used tools by this community, they are imposing themselves over BCs/MCs. As Bernardini and Castagnoli state The term “training” implies that the abilities and competences to be learned are acquirable through practice with the kinds of tools and tasks one will be faced with during one’s future professional career, in an environment that reproduces as closely as possible the future work environment. (2008: 41). Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 72 There seems to be a vicious circle in this respect; the more lecturers use TMs in ‘training’ institutions for translators, the more students will use TMs in their future day to day as professional translators, and TM providers can continue selling them to these institutions reassuring them that they are the tools most used in the translation profession. It is in the hands of translation lecturers to change the course of these events in favour of BCs/MCs by using them more often in their modules. Making students familiar with the tools they will have to use in their future professional practice is very important and no one would want to discuss the importance of TMs in this regard. But there are translation problems related to differences between the pair of languages involved in the translation process that must be dealt with using other kind of tools such as BCs/MCs. According to Bernardini and Castagnoli while the term “training” refers to a situation where students are supposed to work in a simulation of real settings in professional translation, like working under time constraints and using TMs, the term “education” refers to a situation where students would have time for activities and reflections related to cultural and linguistic differences of the pair of languages involved in the translation process that future professional translators will not have time for (cf. 2008:42). For this kind of activities and reflections BCs/MCs are a more suitable tool since they allow the student/researcher to observe and analyse aspects of the language that are beyond the scope of TMs. Lecturers should focus therefore more on ‘educating’ translators, as proposed by Bernardini and Castagnoli (cf. 2008: 39-55), rather than ‘training’ them. On the other hand, translators don’t really have to choose between a TM or a MC for their daily work since these two tools can be used simultaneously complementing each other. Both articles by Bowker/Barlow and Bernardini/Castagnoli quoted above admit that the use of corpora and corpus search facilities among translators in the long term is “likely to depend on their integration with Computer-Aided Translation (CAT) technology” (Bernardini/Castagnoli 2008: 52). In this sense the authors mentioned above plead for a ‘hybrid tool’ (Bowker/Barlow 2003: 20) that combines the strengths of TMs and BCs/MCs. While there is no tool currently that combines all these features, efforts in this direction have been undertaken. MultiTrans, for instance, is the result of implementing Computer Aided Translation Technologies (CAT) with multilingual publishing and document management operations, which allow users to benefit from both systems and has been referred to as “next generation TM” (Gervais 2006: 48). TMs and BCs/MCs also complement each other in the sense that a TM can be converted any time into a parallel corpus to be searched by means of a BC/MC. 2.4. Learning to observe and translate patterns with the help of a parallel corpus: EMEA (European Medicines Agency documents) The fact that corpora can be very helpful during the translational process is something that has been said very often. The question that matters now is: How? Observing patterns in one language and how they have been translated into other languages is a very useful starting point. Sinclair (cf. 2003: xvi-xvii) suggests a seven-stage procedure to read concordance lines for this purpose: 1. Initiate - Search for patterns to the right and left of the node. 2. Interpret - Form a hypothesis that may link these patterns. They may be from the same part of speech or mean something similar. Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 73 Translating, ed. by Allison Beeby, Patricia Rodríguez Inés and Pilar Sánchez-Gijón, 928. Amsterdam/Philadelphia: John Benjamins Publishing Company. Olohan, Maeve and Mona Baker. 2000. Reporting that in translated English: evidence for subconscious processes of explications? Across languages and cultures 1(2): 141158. Sinclair, John. 2003. Reading Concordances. Harlow: Pearson Longman. Varantola, Krista. 2003. Translators and Disposable Corpora. In Corpora in translator education, ed. by Federico Zanettin, Silvia Bernardini and Dominic Stewart, 55-70. Manchester: St. Jerome Publishing. Xiao, Richard. 2009. Well-known and influential corpora. In Corpus linguistics: an international handbook, ed. by Lüdeling, Anke and Merja Kytö, 383-457. Berlin: De Gruyter. Guadalupe Ruiz Yepes Parallel Corpora in7UDQVODWRU redit, número 7, 2011, - 80