Conversation for Textual Case-Based Reasoning

Conversational Case-Based Reasoning (CBR) systems assist their users in formulating queries for case retrieval. Existing textual CBR (TCBR) systems support conversation using only hand-crafted features and indices. However, to be practical, TCBR systems that require conversation should automatically generate their features. This is a difficult problem because TCBR applications routinely involve thousands of interrelated features. In this paper, we explore candidate methodologies to address this problem. We describe domain and human factors issues that must be considered for TCBR conversation. With effective and efficient conversation as our goal, we propose a method for conducting conversation with automatically generated feature vocabularies. We illustrate our approach with examples from a database of air investigation reports and identify underlying problems.