Mining a database of reading mistakes: For what should an automated Reading Tutor listen?

Abstract

Using a machine learning approach to mine a database of over 70,000 oral reading mistakes transcribed by University of Colorado researchers, we generated 225 rules based on graphophonemic context to predict the frequency of the 71 most common decoding errors in mapping graphemes to phonemes. To evaluate their generality, we tested how well they predicted the frequency of the same decoding errors for different readers on different text. We achieved .473 correlation between predicted and actual frequencies, compared to .350 correlation for context-independent versions of the same rules. These rules may help an automated reading tutor listen better to children reading aloud.