{ "paper_id": "1996", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T10:25:34.165815Z" }, "title": "Generalised Language Resources: EURODICAUTOM, SYSTRAN and EURAMIS -a Case Study", "authors": [ { "first": "Jean-Marie", "middle": [], "last": "Leick", "suffix": "", "affiliation": {}, "email": "jean-marie.leick@lux.dg13.cec.be" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "One of the challenges of the EURAMIS project was to find data representation methods capable of covering the language resources of terminology databases, machine translation and translation memories. The importation of EURODICAUTOM terminology data into the SYSTRAN dictionaries gave an indication of the obstacles to be overcome, but it was the exportation of SYSTRAN dictionaries into EURAMIS which revealed the major difficulties. The main problems arose, in the case of EURODICAUTOM, from the lack of strict encoding and, in the case of SYSTRAN, from the algorithmic representation of morphology. Translation memory entries are less problematic provided that logical representation and organic implementation are kept separate. The advantage of common representation is that the various applications access data from different sources using the same mechanisms, thus creating unexpected synergies. The SGML-based approach for logical data description facilitates exchange in TIF-alike format.", "pdf_parse": { "paper_id": "1996", "_pdf_hash": "", "abstract": [ { "text": "One of the challenges of the EURAMIS project was to find data representation methods capable of covering the language resources of terminology databases, machine translation and translation memories. The importation of EURODICAUTOM terminology data into the SYSTRAN dictionaries gave an indication of the obstacles to be overcome, but it was the exportation of SYSTRAN dictionaries into EURAMIS which revealed the major difficulties. The main problems arose, in the case of EURODICAUTOM, from the lack of strict encoding and, in the case of SYSTRAN, from the algorithmic representation of morphology. Translation memory entries are less problematic provided that logical representation and organic implementation are kept separate. The advantage of common representation is that the various applications access data from different sources using the same mechanisms, thus creating unexpected synergies. The SGML-based approach for logical data description facilitates exchange in TIF-alike format.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "At the end of 1992 I took charge of the Multilingual Action Plan (MLAP) at the European Commission, which aimed at lowering language barriers by means of information technology. As the proposals came flooding in for NLP-tools or translation aids, it struck me right from the beginning that every tool, whether it be spell-checker, grammar-checker, term bank or machine translation system, had its own dictionary. While the Commission already had powerful dictionaries, both for its own version of the SYSTRAN machine translation system and for the EURODICAUTOM term bank, these resources could not be used for other tools. Another invaluable, but not easily accessible, language resource was the very large volume of parallel texts in the official languages stored in the archives of the Commission's Translation Service and in the Union's databases (e.g. CELEX). Translators need to have all this data at their fingertips. To this end, a comprehensible toolset working on a common linguistic resources database was designed in the framework of the EURAMIS project (EURopean Advanced Multilingual Information System).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Background", "sec_num": null }, { "text": "This paper discusses the first steps in this direction, which were the integration of the EURODICAUTOM and SYSTRAN dictionaries and their representation within the EURAMIS database.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Background", "sec_num": null }, { "text": "The main achievements, in chronological order, are:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Achievements", "sec_num": null }, { "text": "\u2022 Importation of EURODICAUTOM into SYSTRAN", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Achievements", "sec_num": null }, { "text": "\u2022 Design of the EURAMIS Linguistic Resources Database (LRD)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Achievements", "sec_num": null }, { "text": "\u2022 Exportation of SYSTRAN into the LRD.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Achievements", "sec_num": null }, { "text": "The question of importing EURODICAUTOM terminology data into the SYSTRAN machine translation system had been raised at intervals since machine translation was introduced at the Commission in 1977 and the conclusion had always been -mission impossible!", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Importation of EURODICAUTOM into SYSTRAN.", "sec_num": null }, { "text": "The problem was the fundamental difference in approach between an MT dictionary and a terminology database. MT entries necessarily reflect the most general translation of a term in a large variety of contexts, while terminology entries offer the most precise translation possible in a very specific context. Nevertheless, the wealth of terminology in the EURODICAUTOM base was tempting, and a feasibility study was conducted with Spanish as a source language. The results were so encouraging that they sparked off a revolution.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Importation of EURODICAUTOM into SYSTRAN.", "sec_num": null }, { "text": "While the authors of the feasibility study were optimistic, there were plenty of pessimists to produce counter arguments and examples to show that the results of the feasibility study were biased and at any rate would not apply to other languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Importation of EURODICAUTOM into SYSTRAN.", "sec_num": null }, { "text": "The software specialists feared that performance would deteriorate with the addition of millions of new entries, and the resulting volume explosion on the disks. If panic reigned, this was understandable.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Importation of EURODICAUTOM into SYSTRAN.", "sec_num": null }, { "text": "The nature of Eurodicautom entries presented a real challenge:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 600 000 multilingual entries", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 No formalisation of entries -only guidelines", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 No grammatical information (such as word class or gender)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 No morphological information.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "In short, Eurodicautom entries are meant to be understood by a human reader rather than a machine.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "On the positive side, the project offered:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 Corresponding entries in up to 9 languages", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 Large SYSTRAN dictionaries for cross-checks and for producing languagedependent morphological tables for automatic recognition of word classes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "Hence, we had at our disposal considerable language resources which could be exploited automatically.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "The following approach was adopted:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 Filtering of suitable entries", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 Semi-automatic importation with plausibility checking at every stage", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 Careful fine tuning of strategies.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "Filtering was done on noun groups. Verbs and adverbial groups were excluded.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "Semi-automatic import was achieved by working out probabilities for corresponding entries in the various languages. Nouns could be identified in German (capitalised), verbs in English (preceded by \"to\"), other categories in Latin languages by sorting on endings, etc. The results were checked on sorted lists by the encoders.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "The following diagram shows the process in some detail. The new monoterms were used to produce an MT dictionary, which in turn was used to translate the multiterm expressions. The selection of the best candidate translation via probabilistic methods then generated the new expressions dictionary. The final EDIC-STEM dictionary was generated from information relating to the principal word within the expressions. The development of the whole procedure and the initial import took 180 man/months. This exceeded initial estimates fivefold! (The original workload estimate was 34 man/months), but still resulted in a 600 entries per man day rate. The first two-yearly update only took 20 man-months for 750.000 added entries, giving a rate of 1.700 entries per man-day. It included the development of comparison functions allowing to identify and classify changes to existing entries.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Challenge and the Strategy adopted", "sec_num": null }, { "text": "\u2022 for technical documents: up to 60% or more", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Improvement of translation quality was:", "sec_num": null }, { "text": "\u2022 for other documents: 2-10% or more, depending on the language pair and the type of document Quality enhancements are measured by translating a document twice, once with the EURODICAUTOM dictionaries and once without. An enhancement of 60% means that the number of sentences containing improvements minus the number of sentences containing deteriorations divided by the total number of sentences gives 0,6.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Improvement of translation quality was:", "sec_num": null }, { "text": "The fine tuning provided a better understanding of the Systran system and allowed to streamline coding strategies.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Improvement of translation quality was:", "sec_num": null }, { "text": "The LRD has to cover the spectrum of multilingual entries from paragraph down to stem entries. In order to achieve this the multilingual aligned Text (MAT) -concept was defined as the generalized LRD-entry. Every MAT consists of corresponding monolingual entries referring to linguistic objects with attribute-value pairs describing the details (see diagrams). ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The EURAMIS Linguistic Resources Database (LRD)", "sec_num": null }, { "text": "The structure is described by an SGML construct resulting in an SGML-database.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "EURAMIS LRD design EURAMIS LRD design", "sec_num": null }, { "text": "After the design and initial implementation of the EURAMIS LRD stage, the natural progression was to consider the feasibility of exporting MT dictionaries into the LRD.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "Human nature being as it is, there was again strong opposition. Of course the problems were very real:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "\u2022 Contextual rules in the Systran translation engine", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "\u2022 Morphology, mainly for verbs (paradigms), is defined in Systran in order to optimise retrieval and is dependent on language \u2022 Copyright considerations resulted in proprietary and public data sets", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "\u2022 The data volume explosion by the SGML tags was dramatic", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "\u2022 Character sets in Systran are quite antiquated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "The TIF-DTD was taken into consideration and the LRD-DTD adapted to contain SYSTRAN specific attributes and secrecy provisions. The export procedure from MT was kept separate from the import procedure into the LRD in order to take advantage of linguistic knowledge build-up specific to the LRD and to be able to represent morphology, for example, as an independent factor. Tag compression helped solve the problem of data volume explosion created by SGML-tagging, although the total volume is still estimated at some 20 GB of data for 5 million entries.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "The main work was the definition of the trade-off between representing simple linguistic rules encrypted in the Systran specific coding by standard clauses understandable by linguists. This caused longish discussions and trials in order to isolate specific problems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "The import function for EURAMIS is still under construction. The fine tuning will have to be done after completion of the complete cycle.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Exporting SYSTRAN dictionaries into EURAMIS-LRD.", "sec_num": null }, { "text": "The main benefits of this exercise have been the creation of synergies between developers of machine translation and terminology on one side and combined functions on the other:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "\u2022 Scanning all concerned Eurodicautom entries allowed the correction of entries (spelling-errors, language clean-up, formal variants contradicting the guidelines a.s.o.)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "\u2022 A combined function provides terminology look-up from text in combining the Systran retrieval with Eurodicautom lookup even for target languages not available in Systran.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "\u2022 In using the domain identifiers in terminology entries a content identification function was developed allowing to prescan a document to identify by statistical methods the domain it belongs to.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "\u2022 The use of Eurodicautom entries in Systran translations is of immediate use in machine translation especially for long expressions and lesser developed language pairs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "\u2022 In the future it is planned to use Euramis for entering terminology, so that the MT entries will be available to terminologists.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null }, { "text": "Some of the new products, such as terminology look-up from text, are available today. Others will be available in the EURAMIS system, with the first prototypes currently undergoing tests. We have already achieved a great deal, but there is still more to come!", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": null } ], "back_matter": [], "bib_entries": {}, "ref_entries": { "TABREF0": { "content": "", "num": null, "html": null, "text": "general idea was to separate into monoterm and multiterm expressions and to deal first with monoterms.", "type_str": "table" }, "TABREF1": { "content": "
Importing EURODICAUTOM into SYSTRAN Importing EURODICAUTOM into SYSTRAN
Translation + Correction of misspelled expressionsFR EN
DE
Translate
NFWsSY-Expr.
STEM -EDICEN
FR
DESelect best candidate
translation
Automatic CodingAlternativ in commentCode expression principal word
", "num": null, "html": null, "text": "", "type_str": "table" }, "TABREF2": { "content": "
DG XIII Telecommunications, information market and exploitation of research
GENERALIZED LANGUAGE RESOURCES
Importing EURODICAUTOM into SYSTRAN Importing EURODICAUTOM into SYSTRAN
Systran Dictionaries before and after
importing Eurodicautom data
language pairoriginal entriesnew entries
DE -EN145.000242.341
DE -FR65.000257.492
EN -DE81.000333.644
EN -EL46.000227.649
EN -ES59.000275.524
EN -FR130.000409.027
EN -IT110.000274.584
EN -NL47.000136.817
EN -PT43.000154.163
ES -EN32.000260.865
ES -FR27.000240.132
FR -DE94.000321.372
FR -EN147.000397.956
FR -ES45.000248.802
FR -IT40.000265.753
FR -NL43.000113.738
Total1.154.0004.159.85915
GENERALIZED LANGUAGE RESOURCES
STEM -EDIC
FINAL STEM -EDICEDIC -EXPRESSIONS
14
The final result? More than 3 million entries!
", "num": null, "html": null, "text": "", "type_str": "table" } } } }