Successfully detecting and correcting false friends using channel profiles
Ulrich Reffle, Annette Gotscharek, Christoph Ringlstetter, Klaus U. Schulz · 2008
The detection and correction of false friends - also called real-word errors - is a notoriously difficult problem. On realistic data, the break-even point for automatic correction so far could not be reached: the number of additional infelicitous corrections outnumbered the useful corrections. We present a new approach where we first compute a profile of the error channel for the given text. During the correction process, the profile helps to restrict attention to a small set of suspicious lexical tokens of the input text where it is plausible to assume that the token represents a false friend. Using a conventional word trigram statistics for disambiguation we obtain a correction method that can be successfully applied to unrestricted text. In experiments for OCR documents, we show significant accuracy gains by fully automatic correction of false friends.