Corpus-based Error Detection in a Multilingual Medical Thesaurus

Cross-language document retrieval systems require support by some kind of multilingual thesaurus for semantically indexing documents in different languages. The peculiarities of the medical sublanguage, together with the subjectivism of lexicographers' choices, complicates the thesaurus construction process. It furthermore requires a high degree of communication and interaction between the lexicographers involved. In order to detect errors, a systematic procedure is therefore necessary. We here describe a method which supports the maintenance of the multilingual medical subword repository of the MorphoSaurus system which assigns language-independent semantic identifiers to medical texts. Based on the assumption that the distribution of these semantic identifiers should be similar whenever comparing closely related texts in different languages, our approach identifies those semantic identifiers that vary most in distribution comparing language pairs. The revision of these identifiers and the lexical items related to them revealed multiple errors which were subsequently classified and fixed by the lexicographers. The overall quality improvement of the thesaurus was finally measured using the OHSUMED IR benchmark, resulting in a significant improvement of the retrieval quality for one of the languages tested.