Learning from Textbook Knowledge: A Case Study

Conclusions To summarize, we have identiied a problem that arises in learning from the knowledge in textbooks: the problem of learning from knowledge including omissions and inconsistencies that are clariied by illustrative examples. This learning problem is solvable by a technique that we call analogical abductive explanation based learning (ANA-EBL). ANA-EBL actually solves the more general problem of learning from understanding examples using an approximate theory. It is an explanation based learning technique that combines techniques used to choose between multiple inconsistent explanations with explanation based analogical reasoning techniques along the lines of Huhns and Acosta, 1987; Kedar-Cabelli, 1987]. Learning from textbook knowledge is a hard problem ; even if the natural language problems are nessed by manually translating a theory into logic, the problem of correcting and completing the resulting theory is diicult. The major contribution of this paper is isolation of some subproblems involved in the general problem of learning from textbooks, and presentation of techniques that address these subproblems. The techniques developed to solve this problem, however, are of independent interest. ANA-EBL is a theory specialization technique, like the techniques however, ANA-EBL works even in situations in which the original theory is re-lational and/or generates multiple inconsistent explanations , and in which the target theory is disjunc-tive. Experimental results indicate that the technique is eeective on randomly selected examples, as well as on well-chosen \textbook" examples. ANA-EBL also builds on techniques described in Pazzani, 1988] and Fawcett, 1989] for solving the multiple explanation problem but extends these results by rst, incorporating analogical reasoning techniques similar in avor to those used in Huhns and Acosta, 1987; Kedar-Cabelli, 1987]; and second, giving a precise way of weighting the complexity of an explanation and the number of observations that it covers, and justifying this heuristic with a formal analysis. ANA-EBL is far from a complete solution to the problem of learning from textbook knowledge; however , we feel that it addresses at least some of the issues that must be confronted. An important topic for further research would be to integrate ANA-EBL with an automatic text understanding system. This would address the major shortcoming of our evaluation of ANA-EBL: in the process of manually translating a text into Horn clause logic, representational choices that aaect learning inevitably must be made. Because of this, the skeptical reader may wish to view this work as learning from a synthetic theory that is …