Semantic transference for enriching multilingual biomedical knowledge resources

Abstract

Biomedical knowledge resources (KRs) are mainly expressed in English, and many applications using them suffer from the scarcity of knowledge in non- English languages. The goal of the present work is to take maximum profit from existing multilingual biomedical KRs lexicons to enrich their non-English counterparts. We propose to combine different automatic methods to gener- ate pair-wise language alignments. More specifically, we use two well-known translation methods (GIZA++ and Moses), and we propose a new ad-hoc method specially devised for multilingual KRs. Then, resulting alignments are used to transfer semantics between KRs across their languages. Transfer- ence quality is ensured by checking the semantic coherence of the generated alignments. Experiments have been carried out over the Spanish, French and German UMLS Metathesaurus counterparts. As a result, the enriched Span- ish KR can grow up to 1,514,217 concepts (originally 286,659), the French KR up to 1,104,968 concepts (originally 83,119), and the German KR up to 1,136,020 concepts (originally 86,842).