Motivation: Millions of protein sequences currently being deposited to sequence databanks will never be annotated manually. Similarity-based annotation generated by automatic software pipelines unavoidably contains spurious assignments due to the imperfection of bioinformatics methods. Examples of such annotation errors include over- and underpredictions caused by the use of fixed recognition thresholds and incorrect annotations caused by transitivity based information transfer to unrelated proteins or transfer of errors already accumulated in databases. One of the most difficult and timely challenges in bioinformatics is the development of intelligent systems aimed at improving the quality of automatically generated annotation. A possible approach to this problem is to detect anomalies in annotation items based on association rule mining. Results: We present the first large-scale analysis of association rules derived from two large protein annotation databases - Swiss-Prot and PEDANT - and reveal novel, previously unknown tendencies of rule strength distributions. Most of the rules are either very strong or very weak, with rules in the medium strength range being relatively infrequent. Based on dynamics of error correction in subsequent Swiss-Prot releases and on our own manual analysis we demonstrate that exceptions from strong rules are, indeed, significantly enriched in annotation errors and can be used to automatically flag them. We identify different strength dependencies of rules derived from different fields in Swiss-Prot. A compositional breakdown of association rules generated from PEDANT in terms of their constituent items indicates that most of the errors that can be corrected are related to gene functional roles. Swiss-Prot errors are usually caused by under-annotation owing to its conservative approach, whereas automatically generated PEDANT annotation suffers from over-annotation.