Users Are Always Right ... Even When They Are Wrong: Making Knowledge Representation Useful and Usable

Users’ problems in developing applications rarely factor easily so as match the solutions provided by the Knowledge Representation community. If KR is to be applied successfully, the gap between users’ “problem space” and the KR “solution space” must be addressed. Two approaches are suggested: “Intermediate representations” at a higher level of abstraction than current KR languages and a set of application oriented questions to expose how different systems address common user questions. The results are likely to be hybrid systems including heuristics and ‘non-standard’ reasoning. The Problem Space and the Solution Space What counts as successful knowledge representation? At first glance the answer seems simple: effective use in knowledge based applications that could not, or would not, have been built or maintained otherwise. Applications are built to solve problems – and ultimately problems ‘belong’ to users. At one level, we appear to be in a golden age of Knowledge Representation with the explosion of the Semantic Web, and widespread recognition of RDF, OWL and related standards. On another level, there remains the same gap between the practitioners and the users pointed out by Doyle and Patil in the early 1990s (Doyle and Patil 1991) – the formalisms that are tractable and well understood do not match users’ needs. Yet most of the Knowledge Representation community tends to be focused on technical arguments about different formalisms or solutions – DLs, OWL, Rules, CLP, Belief Nets, Machine Learning, etc. – rather than users’ problems. Users are the experts in their problems – even when they present them as (possibly naïve) solutions. Users’ need an improved supply chain manager, an adaptable clinical data capture tool, a repository for biological data, or whatever. Users define the “problem space”. More often than not will find a way to achieve something that approximates their ends, however ill founded the technologists think it in terms of the “solution space”. Compilation copyright © 2006, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. Not infrequently users want “an answer” rather than “the best answer” – an improvement to the status quo rather than perfection. Furthermore, the limitations users accept may not reflect technologists’ values. For example, a user may value predictability or expressiveness over completeness or even decidability. There is an analogy with the well-known quote from John Tukey (Tukey 1962). For a user it may be: ...far better to have an approximate answer to the right question ... than the exact answer to the wrong question... None of this is original. It is the stuff of user-centred design textbooks and a key part of knowledge representation paradigms such as CommonKADS (Schreiber 2000). The issues date back to the arguments between the ‘neats’ and ‘scruffies’ of early AI (Sloman 1990) now extended to philosophical ontology domain (Goble and Wroe 2004). The Challenge of the Problem Space Reformulating users’ problems in terms of available solutions is at the core of many other disciplines, including the author’s field of Healthand BioInformatics. Too often the KR community appears to put the burden on users to reformulate the problem in terms of the KR community’s criteria for judging solutions. Add to this is the fact that logic is daunting to many users, even users sophisticated in other forms of mathematics. Add further to this that KR language features are more often chosen for computational tractability and logical elegance than usability or continuity. Add still further serious discontinuities – the fact that newer formalisms do not always answer questions that were the raison d’être for their predecessors. Then it is wonder that there is confusion and skepticism as we try to build standards around more sophisticated KR languages such as OWL. How can the KR community address these problems? Clearly not by abandoning its high standards for precision in formulating the solution space. However, that focus needs to be supplemented by renewed attention to the problem space both theoretically and empirically. A userand application-oriented vocabulary is needed for user issues. The Semantic Web Best Practice and