Extracting semantic relations

Henny Klein (Rijksuniversiteit Groningen)
Lolkje de Jong-van den Berg (Rijksuniversiteit Groningen)
Rein Vos (Rijksuniversiteit Groningen)

Extracting terms that are engaged in a specific relationship is an 
important objective in IR research: the subject of a paper may be 
characterized by its (key) words, but to decide whether a paper is of 
interest it is often important to know on which relations between those 
terms the text focusses. 

As our group is interested in adverse drug reactions (ADRs), we tested three approaches to find the combinations of a drug and an associated ADR in medical texts. In these experiments, we made use of an NLP program (MetaMap, AR Aronson c.s.) to map the text to medical concepts represented in the UMLS Metathesaurus.

We conclude that to achieve an acceptable performance in extracting drug -ADR combinations, it is necessary to check on the concepts as well as their relation. However, due to characteristics of the drug - ADR relation, the variety of expressions used to refer to it seems to be unlimited. In these aspects our subject differs from those in other relation extraction experiments in the medical field, where the sought- for relation is either inferred from the collocation of concepts (drug - dosage) or expressed by one standard term (x binds y).