This talk adresses the challenges of semi-automatic information-integration and abstraction in real-world domains. While there is no lack of approaches to this problem, there are often severe restrictions limiting the applicability in challenging domains, for example concerning limited expressivity (Bayes Nets) or problems in handling imperfect date (many logic-based approaches).

We propose the use of abductive reasoning to handle two prominent qualities of information imperfection, namely incompleteness (lack of information), and uncertainty (degree of belief in the truth oof some piece of information). Abductive inference schemes naturally handle missing data by allowing for assumptions, and weighting based on probabilities allows to integrate uncertain information both on instance and schema level. I conclude by presenting ideas on additional information-teoretic and cost criteria for guiding the abductive process in the context of situation recognition, and for triggering information-gathering activities.