In this paper we describe how we applied a general- purpose machine translation tool for translating biomedical thesauri. We used corresponding terms in parallel corpora to check the validity of the translations. The advantage of this approach is that a single corresponding set of terms can be verified where techniques to retrieve translations from a parallel corpus do not exploit the knowledge contained in current state of the art machine translation software.
second translated thesaurus only based on the existing non-English terms in the UMLS. This UMLS-translated thesaurus was included as a baseline thesaurus to assess the improvement that could be obtained with the machine translation. We applied this approach for two languages, Dutch and German.
We used two parallel corpora: multilingual drug labels from the EMEA corpus, and bilingual titles of scientific abstracts from Medline. 3
We indexed the parallel corpora with the English thesaurus, the UMLS-translated thesaurus, and the machine-translated thesaurus. We differentiated for the different semantic types and the different parallel corpora we have (for German restricted to the drugs labels from EMEA and MedLine titles The results for Dutch, German, and French are show in table 1. Each table shows the results for finding concepts using only the manual translation as contained in the UMLS and the results when the machine translated terms are added. We provide not only the figures for the translated terms that have correspondences in English and the translation language (BOTH), but also terms that have only be found in English (ENGLISH) or only in the translation language (DUTCH, GERMAN, or FRENCH). The tables show the results for the EMEA drug label parallel corpus and for the Medline titles.
Tabel 1. Results for the Dutch, German, and French EMEA drug labels and Medline titles.
ENTRY
OBJC
GEOG
CHEM
DEVI
PHEN
DISO
ANAT
LIVB
PHYS
TOTAL ENGLISH BOTH FRENCH ENGLISH BOTH FRENCH ENGLISH BOTH FRENCH ENGLISH BOTH Language Dutch German French 102 Tabel 2. Overall statistics for the different languages and corpora.
EMEA
Medline Original
Translated
Original
Translated 4
The results show that machine translation can help to enrich a thesaurus. Compared with the manual UMLS-translated thesaurus, the number of terms in the machinetranslated thesaurus that are found in the parallel corpora, doubles for German, French and almost triples for Dutch when only considering concepts that have been found in English. This is consistent for both parallel corpora included in this evaluation. The German and French manual translations are more extensive than the Dutch one, which likely explains the difference in number of extra terms found. The increase is largest for some semantic groups that have hardly been translated (objects, devices, and chemicals). If one also looks at terms that have only been found in the translated corpus and not in the original English corpus an additional set of translated terms can be found. We will also evaluate this set of terms only found in the translated corpus for correctness. We will extend this evaluation to include Spanish as well.
van der Plas L, Tiedemann J. Finding medical term variations using parallel corpora and distributional similarity. In: Proc 6th Workshop on Ontologies and Lexical Resources, 2010.
Och FJ, Ney H. A systematic comparison of various statistical alignment models. Comp Ling 2003;29:19-51.
Schulz S. et al. Machine vs. human translation of SNOMED CT terms. Currently under review for MEDINFO 2013 https://developers.google.com/translate/ Bodenreider and O, McCray, AT. Exploring semantic groups through visual approaches, Journal of Biomedical Informatics 36(6):414–432, 2003.