Interleaving Reasoning and Selection with Knowledge Summarization

Abstract

When the physical space and the cyber space are linked by human, Cyber-Physical Society (CPS) has emerged and produced many challenges. Among which, the challenge of fast growing data and knowledge both from the physical space and the cyber space has become a crucial issue. Scalability becomes a big barrier in data processing (more specifically, search and reasoning). Traditional knowledge processing methods aim at providing users complete results in rational time, which is not applicable when it comes to very large-scale data. While in the context of Web and large-scale data, users' expectations are not always receiving complete results, instead, they may prefer to get some incomplete subset of the results compared to waiting for a long time. With this spirit, an approach named Interleaving Reasoning and Selection with Knowledge Summarization (IRSKS) is developed. This approach supports incomplete reasoning and heuristic search based on knowledge summarization. It can be divided into two phases: the off-line and the on-line processing. The off-line processing includes partitioning and summarization, and provides the basis for heuristic search. Partitioning makes one large-scale dataset become many small subsets (chunks). Summarization produces summaries that contain heuristic information (such as the location and major topics of each chunk) to build a bridge between the searching target and the partitioned large-scale dataset. Through the cues provided by summaries, a best search path can be found to locate the searching target. Along with the search path, the on-line processing includes interleaving reasoning and selection, which compose a dynamic searching process and support anytime behavior. In this way, the search space is greatly reduced and close to the searching target so that a good trade-off is achieved between the time and the quality of a query. Based on this approach, a prototype system named Knowledge Intensive Summarization System (KISS) has been developed and the evaluation with the KISS system on the PubMed dataset indicates that the proposed method is potentially effective for processing large-scale semantic data in the Cyber-Physical Society.