MANTRA · Multilingual Annotation of Named Entities and Terminology Resources Acquisition

This project will provide multilingual terminologies and semantically annotated multilingual documents, e.g., patent texts, to improve the accessibility of scientific information from multilingual documents. The two SME partners will use these resources to improve the quality and functionality of their product offerings, viz. delivering multilingual search and text mining engines based on multilingual terminologies. Both SMEs will market these solutions to their customer base.The MANTRA project capitalizes on parallel document corpora from which translational correspondences will be computed by the use of different alignment methods. Fortunately, the biomedical domain, the application scenario of MANTRA, offers a rich variety of such parallel corpora. We will exploit these multilingual document sets to harvest terms and concept representations in different languages in order to augment currently available terminological resources such as the Medical Subject Headings (MeSH).The project partners will collaboratively build two types of resources: automatically enhanced multilingual terminologies and semantically annotated multilingual documents. The novelty of the latter resource derives from the fact that we solicit and orchestrate community efforts for building up these annotated resources, a procedure that has already been proven successful for the semantic enrichment of large-scale biomedical document corpora (CALBC project) which was executed by the project partners. The novelty of the first comes from a new combination of existing technologies in the area of statistical machine translation, named entity tagging and terminological resources. We start from statistically aligned, parallel documents on which named entity taggers are run to produce highly diverse semantic (named entity) annotations. These annotations signal concept mentions in the text which can then be linked to corresponding entries in relevant biomedical ontologies (from the UMLS, OBO or BioPortal umbrellas), and, in addition, provide the corresponding concept identifiers. Parallel named entity occurrences lacking links to the chosen ontologies can be considered as putative translation equivalents. Validated putative translation equivalents can then be used to enhance already given monolingual terminological resources. Both types of resources will be made available to the public for translation purposes and for search in and text mining from multilingual documents.