Detecting Malapropisms Using Measures of Contextual Fitness

While detecting simple language errors (e.g. misspellings, number agreement, etc.) is nowadays standard functionality in all but the simplest text-editors, other more complicated language errors might go unnoticed. A difficult case are errors that come in the disguise of a valid word that fits syntactically into the sentence. We use the Wikipedia revision history to extract a dataset with such errors in their context. We show that the new dataset provides a more realistic picture of the performance of contextual fitness measures. The achieved error detection quality is generally sufficient for competent language users who are willing to accept a certain level of false alarms, but might be problematic for non-native writers who accept all suggestions made by the systems. We make the full experimental framework publicly available which will allow other scientists to reproduce our experiments and to conduct follow-up experiments.