bloodPressureDiastolic and hasValue:95. bloodPressureSystolic and hasValue:150. bloodPressureDiastolic and hasValue:90. bloodPressureSystolic and hasValue:130. tities. Thus, hce2 will still be known to be an instance of them, and the other maps concept names to the set of We can now dene a set of individuals to be indexed ing concepts (e.g., the concepts specializing patient), and individuals they directly instantiate, i.e. those which are examSomeBpSysAbnorm, but the observations and their shaded part in Fig. 1), but to keep all the information store the relation which associates each indexing concept with the individuals it instantiates. This relation can efthat was derived from observations concerning other enthe associated cardinalities. values from which this was derived will become unknown. remove all the information concerning observations (the not instances of any subconcept. It is also useful to store These two properties make it possible, for example, to cien tly be stored in two hashtables: one maps individ(for example the set of patients), choose a set of indexual names to the set of most specic concepts describing for each individual the set of concepts it is (and is not) an instance of. * patient anything
* examination examAllBpSysNorm patSomeBpAbnorm examAllBpDiaNorm patAllBpNorm examAllBpNorm examSomeBpDiaAbnorm patAllBpDiaNorm examSomeBpAbnorm observation * * bloodPressure * * Norm bPSystolic * * bPDiastolic Abnorm patSomeBpSysAbnorm subsumed by instantiates filled by minological systems. The indexing elements are potenfor computing subsumption are not disastrous for indexing: they will simply result in a less informed, suboptitially complex descriptions logically related by subsumpmal index.
This semantic indexing mechanism is crucially depention and disjointness. Note that incomplete algorithms dent on reasoning with descriptions as provided by terbetween, some partial information. each candidate instance the original description must be accessed and explicitly tested against the query. After of indexed instances. This means no information at all puting intersections and unions, etc. In case the query is Otherwise there is a remaining set of candidates, the (and cardinality) can be computed. ity. If we ask for a concept which is totally unrelated to this has been done, the user can choose to declare the ther true or false. In this case the index alone does not contain enough information to determine the extension a combination of indexing concepts, its exact extension dense at that particular point in the semantic space. ing indexing concepts. If we ask for a concept which is no subsumed, and no disjoint ones, we will get a lower The second phase additionally utilizes the actual exfrom the index. Typically, one should get something in bound of 0 and an upper bound equal to the number This generally results in much better cardinality estiindividuals for which the query is not known to be eitensions of indexing concepts also stored in the index. of the query, and the third phase must be entered. For mates at the cost of having to load the instances, comquery as a new indexing concept, making the index more equivalent to an indexing one, we get the exact cardinalexisting indexing concepts, i.e. there are no subsuming, male