UNCORRECTED DRAFT . For the final version , see Automated Building of

For most languages, including Polish, big error corpora are lacking. Traditional error corpora are collected and annotated by linguists, and the process is manual or only slightly automated. The task is therefore tedious and costly, and the results represent linguists’ knowledge about correct usage. This requires additional work to avoid theory-laden distortion of data. In this paper, I will show how to automatically develop error corpora by using revision histories of documents. The idea is based on a hypothesis that most frequent minor edits in documents represent corrections of typos, slips of the tongue, grammar, usage and style mistakes. This hypothesis has been confirmed by frequency analysis of the revision history of articles in the Polish Wikipedia. Partial results of the analysis and perspectives for integrating the error corpus with the Polish National Corpus will be presented.