Detecting annotation errors in a corpus by induction of syntactic patterns

Abstract

This paper brings a new method for acquisition of syntactic patterns capable of detecting errors in annotated corpora. These patterns are acquired semi-automatically, by means of an inductive logic programming (relational data mining) system followed by a human expert supervision. The patterns acquired have been used for automatic detection and subsequent manual correction of the annotation errors found in DESAM, a morphologically annotated corpus of written Czech. Preliminary results show efficiency of the method: more than 7000 annotation errors in the corpus have been successfully detected and corrected so far.