From: Patrick Haluptzok (patrickh_at_windows.microsoft.com)
Date: Wed Oct 22 2003 - 07:49:46 PDT
CSE 573 Paper Review #2 - Reviewed by Patrick Haluptzok Oct 21, 2003
Two Theses of Knowledge Representation
Language Restrictions, Taxonomic Classification,
and the Utility of Representation Services
by Jon Doyle and Ramesh S. Patil
Summary:
The authors argue that omitting constructs from a language in a knowledge representation system to optimize for classification speed is misguided because the resulting language loses important functionality, alternative approaches to have fast classification exist and a system's total utility is more that just its ability to do classification polynomial time.
Main Points:
Removing constructs from a representation language causes it to lose critical expressive power. The describe how KL works and then show a list of first order logic sentences that would commonly be encountered in representing information for a model that can't be represented in KL. The paper explains that people encoding information must work around these deficiencies in a way compatible with how users of the system will work around these deficiencies - conventions on how to represent concepts and relations that KL doesn't directly support must be established. As a result they claim the resulting KL systems have many relations added as primitives, limiting the functionality of the system.
Knowledge representation systems utility is more than just guaranteeing polynomial time. They describe how the restricted language can't represent all concepts and thus are unable to perform queries and classifications a first order logic representation would allow. So although the system is guaranteed fast on classification, important functionality is missing from the system. They go on to describe the extra costs that knowledge encoding and end users of the system incur trying to work around the limitations. They argue providing a more general representation is preferable because the overall value of the system is higher and other compromises can be employed to get faster classification when it is needed.
Main Flaws:
The paper lacked sufficient case studies of knowledge representation systems showing how the overall utility was better when using more expressive logic languages for the knowledge base. The paper was very critical of KL, but didn't come off balanced. It lacked good concrete comparisons of systems using KL and more expressive representations.
The paper was long to present the information in it. The paper complained about KL quite a bit but didn't convince me that their proposals would yield a system still as useful in the end. It had qualitative recommendations that weren't backed up quantitatively.
Future Research:
I'd like to see how the performance of a KL system compared with a more full representation for performing classification as the database scaled in size, because that is a critical measure for many applications. A comparison of the time to encode the database in KL versus a more expressive representation would be important; also a comparison of the time for users to construct queries in each language; both of these measures relate to the overall utility of a system.
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 07:49:56 PDT