In this paper, we introduce a partially automated method to generate qualified answers at multiple abstraction levels for database queries. We examine the issues involving data summarization by Attribute-Oriented Induction (AOI) on large databases using fuzzy concept hierarchies. Because a node may have many abstracts, the fuzzy hierarchies become more complex and vaguer than crisp ones. Therefore, we cannot use exactly the original AOI algorithm with crisp hierarchies, applied for fuzzy hierarchies, to get interesting answers. The main contribution of this paper is that we propose a new approach to refine fuzzy hierarchies and evaluate tuple-terminal conditions to reduce noisy tuples. The foundations of our approach are the generalization hierarchy and a new method to estimate tuple quality. We implemented the algorithm in our knowledge discovery system and the experimental results show that the approach is efficient and suitable for knowledge discovery in large databases.