in projects associated with the employee. Whereas with self-published documents there is an explicit linkage between the documents and employee (they are indexed by employee number on the MII), with documents that mention employees, this linkage must be derived from the underlying text. In particular, an information extraction tool1 tags proper names in text and then statistically measures how strongly these names are associated with specific topics. Because each source of information alone is not sufficient to determine if an employee is an "expert" in a particular topic, ExpertFinder relies on the combination of evidence from many sources.
The original goal of ExpertFinder was to place a user within one phone call of an expert. That is, even if the persons listed as the result of an ExpertFinder query weren't the experts, they would be able to provide the name of someone who was. Happily, actual "experts" are typically listed as the top three or four candidates. The likelihood of randomly selecting a correct expert is the total number of experts divided by total corporate staff (4500) so there is often significantly less than a one-percent chance of finding any experts.
Table 1 contrasts the performance of ten technical human resource managers, professionals at finding experts, with ExpertFinder for the task of identifying the top five corporate experts in speciality areas listed in the table. The first column in Table 1 shows the degree of inter-subject variability in reporting experts (measuring percentage of agreement of first, second, and third of five experts). Columns 2 and 3 show results for precision, the degree to which a staff found by ExpertFinder is considered expert by humans, and recall, the degree to which apriori human-designated experts are found by ExpertFinder.____________________