Automation of dictation exercises. A working combination of CALL and NLP.

This article is in the context of the Computer-Assisted Language Learning (CALL) frame- work, and addresses more specifically the automation of dictation exercises. It presents a method for correcting learners' copies. Based around Natural Language Processing (NLP) tools, this method is original in two respects. First, it exploits the composition of finite- state machines, to both detect and delimit the errors. Second, it uses automatic morpho- syntactic analysis of the original dictation, which makes it easier to produce superficial and in-depth linguistic feedback. The system has been evaluated on a corpus of 115 copies including 1,532 copy errors. The accuracy of the error detection is 99%. The superficial feedback is 97.2% correct, the in-depth feedback 96%, and the morpho-syntactic analysis 87.7%.