Automatically Acquiring Structured Case Representations: The SMART Way

Abstract

Acquiring case representations from textual sources remains an interesting challenge for CBR research. Approaches based on methods in information retrieval require large amounts of data and typically result in knowledge-poor representations. The costs become prohibitive if an expert is engaged to manually craft cases or hand tag documents for learning. Thus there is a need for tools that automatically create knowledge-rich case representations from textual sources without the need to access large volumes of tagged data. Hierarchically structured case representations allow for comparison at different levels of specificity thus resulting in more effective retrieval than can be achieved with a fiat structure. In this paper, we present a novel method for automatically creating, hierarchically structured, knowledge-rich cases from textual reports in the Smart-House domain. Our system, SMART, uses a set of anchors to highlight key phrases in the reports. The key phrases are then used to learn a hierarchically structured case representation onto which reports are mapped to create the corresponding structured cases. SMART does not require large sets of tagged data for learning, and the concepts in the case representation are interpretable, allowing for expert refinement of knowledge.