Compositional Modeling for Computer-Based Tutoring of Prediction Tasks

摘要

Our previous work has demonstrated the utility of computer based advisorysystems, such as expert systems and tutoring systems. We have developed methods for automatically answering a wide assortment of questions, even questions that were not anticipated when the advisory system was built. We have evaluated our question answering methods by comparing their explanations with those written by human experts. Using a 'Turing test' experimental design. The results were very encouraging: a separate panel of human experts graded our machine generated explanations only slightly lower than the human generated ones. Despite the effectiveness of computer based advisory systems, a major obstacle prevents their widespread development and deployment: the knowledge base underlying each system is extremely difficult to build. Although we have built several knowledge bases, in such diverse domains as legal reasoning and biology, each one was built 'from scratch', with little transfer from other knowledge bases. Despite our considerable experience building such systems, our largest knowledge base required about ten man-years of sustained effort. Our research during the past year has focused on this problem. We have developed methods for building knowledge bases from reusable components, analogous to the way that large 'object oriented' software systems are built.