Expert Knowledge and Embedded Knowledge: Or Why Long Rambling Class Descriptions are Useful

Abstract

In many natural resource inventories class descriptions have atrophied to little more than simple labels or ciphers; the data producer expects the data user to share a common understanding of the way the world works and how it should be characterized (that is the producer implicitly assumes that their epistemology, ontology and semantics are universal). Activities like the UK e-science programme and the EU INSPIRE initiative mean that it is increasingly difficult for the producer to anticipate who the users of the data are going to be. It is increasingly less likely that producer and user share a common understanding and the interaction between them necessary to clarify any inconsistencies has been reduced. There are still some cases where the data producer provides more than a class label making it possible for a user unfamiliar with the semantics and ontology of the producer to process the text and assess the relationship between classes and between classifications. In this paper we apply computer characterization to the textual descriptions of two land cover maps, LCMGB (land cover map of Great Britain produced in 1990) and LCM2000 (land cover map 2000). Statistical analysis of the text is used to parameterize a look-up table and to evaluate the consistency of the two classification schemes. The results show that automatic processing of the text generates similar relations between classes as that produced by human experts. It also showed that the automatically generated relationships were as useful as the expert derived relationships in identifying change.