Learning User-Defined, Domain-Specific Relations: A Situated Case Study and Evaluation in Plant Science

Although methods exist to identify well-defined relations, such as is_a or part_of, existing tools rarely support a user who wants to define new, domain-specific relations. We conducted a situated case study in plant science and introduce four new domain-specific relations that are of interest to domain scientists but have not been explored in information science. Results show that precision varies between relations and ranges from 0.73 to 0.91 for the manufacturer location category, 0.89 and 0.93 for the seed donor-bank relation, 0.29 and 0.67 for the seed origin location, and 0.32 and 0.77 for the field experiment location. The manufacturer location category recall varies from 0.91 to 0.94, the seed bank-donor location recall ranges between 0.93 and 1, the seed origin relation from 0.33 to 0.82 while the field experiment location from 0.67 to 0.83 depending on the classifier and using a combination of lexical and syntactic features in the background.