To prevent the disclosure of privacy, it requires preserving the anonymity of sensitive attributes in data sharing. The attribute values on quasi-identifiers often have to be generalized before data sharing to avoid linking attack, and thus to achieve the anonymity in data sharing. Data generalization increases the uncertainty of attribute values, and results in the loss of information to some extent. Traditional data generalization is often based on the predefined hierarchy, which causes over-generalization and too much unnecessary information loss. In this paper, the attributes in a quasi-identifier are classified into two categories, ordered attributes and unordered attributes. More flexible strategies for data generalization are proposed for them, respectively. At the same time, the loss of information is defined quantitatively based on the change of uncertainty of attribute values during data generalization. Furthermore, data anonymization is modeled by a clustering problem with special constraints. A clustering-based approach, called L-clustering, is presented for the l-diversity model. L-clustering can meet the requirement of preserving anonymity of sensitive attributes in data sharing, and reduce greatly the amount of information loss resulting from data generalization for implementing data anonymization.