The Role of Coherence in Constructing and Evaluating Abductive Explanations

摘要: I lying much of human intelligent activities, including together with background knowledge, logically entails text understanding, plan recognition, disease diagno- a set of observations [CM85]. sis, and physical device diagnosis. In this paper, we We have built a language understanding system describe some problems encountered using abduction called ACCEL (Abductive Construction of Causal Exto understand text, and present some solutions to over- planations for Language) that is capable of constructcome these problems. The solutions we propose center ing deep, causal explanations for natural language text around the use of a different criterion, called explana-(both narrative and expository text) through the use tory coherence, as the primary measure to evaluate the of abduction. ACCEL includes a generic abductive inquality of an explanation. In addition, explanatory co- ference procedure, which computes abductive proofs herence plays an important role in the construction of by backward-chaining on the input observations using explanations, both in determining the appropriate level Horn-clause axioms in the knowledge base. The abof specificity of a preferred explanation, and in guiding ductive procedure has the choice of making a subgoal the heuristic search to efficiently compute explanations in a partial proof as an assumption, if it is consistent to of sufficiently high quality. do so. An abductive proof represents an explanation, or an interpretation of the input sentences.