Case-Based Reasoning Research Summaries

CBR-IR

Projects in the CBR-IR Group explore various uses of CBR on problems of information retrieval. In the BankXX system, we explored the use of an architecture based on heuristic search to discover, peruse, and retrieve information from a highly interconnected network of information, such as found in the literature from domains such as law and medicine. In ongoing work, we are exploring hybrid approaches that use highly articulated CBR models of relevance to drive INQUERY-style fulltext IR. Currently, we are using CBR to seed IR: that is, we use results of a CBR analysis to cause INQUERY's relevance feedback to automatically generate a query, which INQUERY then acts on in its usual manner. Preliminary results show signficant, robust improvement over the use of IR alone. Case-based planning to formulate retrieval plans, based on past experience and the user's current information need and factual context, is another approach for combining CBR and IR; this was explored in the FRANK system.

Publications

BankXX

BankXX is the first CBR system using the framework of heuristic best-first search to guide retrieval of cases and other pertinent knowledge for adversarial argument. It searches through a case-domain-graph whose nodes represent key items of legal knowledge, such as cases, legal theories, stereotypical factual scenarios. Search is guided by one of three heuristic evaluation functions-called node-type, argument-piece, argument-dimensions. Each takes into consideration how the information in an (open) node can contribute to the evolving argument being built up as BankXX searches.

Publications

FRANK

FRANK is a blackboard-based architecture to create diagnostic reports tailored to the user's prescribed goals and specified report type. This mixed paradigm system, incorporating CBR, rule-based and planning components, dynamically modified its retrieval strategies and queries with feedback from the system's previous success or failure in retrieving cases that support a user's rhetorical and pragmatic goals for the report.

Publications

OFF-BROADWAY

A case retrieval and classification system that learns a set of distinguished cases (actual prototypes) that have demonstrated classification power. The system maintains a population of sets of potentially prototypical cases; each set is evaluated by its classification accuracy. A genetic algorithm is used to search the space of prototype sets ("spanning sets" in the language of ?I.2.B) for one of superior accuracy.

Publications

BROADWAY

This system represents cases as blackboard knowledge sources whose preconditions invoke local similarity functions that apply only within a closed neighborhood of each in the space of cases. Broadway provided a new dynamic model of case retrieval in which individual cases are assigned their own similarity metrics, so that retrieval varies according to the type of case being retrieved.

Publications

CABOT

This project in case-based search developed an inductive learning algorithm that adjusts its retrieval and adaptation mechanisms. CABOT demonstrated a 50% reduction in the number of stored cases and an increased level of performance in a game-playing application.

Publications

CABARET/ID5

This CBR/ML project investigated the selection of training instances using characterizations of cases (e.g., most on-point, trumping, best) employed in our other CBR systems. In evaluating several different collections of training instances, we found that some (e.g. sets using "trumping" cases) were particularly successful in inducing decision trees generated by Utgoff's ID5.

Publications

CABARET

CABARET is the first mixed paradigm (CBR-RBR) hybrid. It combines HYPO-style CBR with standard (forward and backward) rule-based reasoning through the use of heuristic control rules. The control rules embody theory of statutory argument composed of argument stances, moves and primitives. CABARET-style arguments address key issues in rule interpretation, including open-textured predicates and rules with emergent exceptions and unstated prerequisites.

Publications

HYPO is the first precedent-based CBR system. Its novel contributions include: dimensions, a kind of index used for case comparison as well as case retrieval; computational definitions of important types of cases for argument including most on-point, best, and trumping cases; definition of a partial ordering (the claim lattice) of relevant cases according to how on-point they are; 3-ply arguments built up by analogizing and distinguishing cases; and hypothetical generation, "what if" adaptations of cases to test an argument.

Publications