Chunking: an unsupervised method to find errors in text

Abstract

We describe a method to use a chunker for grammar checking. Once a chunker is available the method is fully unsupervised, only unannotated text is required for training. The method is very simple, compare the output of the chunker on new texts to the output on known correct text. Rare chunk sequences that occur in the new texts are reported as suspected errors. By automatically modifying the chunk set to be more detailed for common verbs or prepositions more error types can be detected. The method is evaluated on Swedish texts from a few different genres. Our method can be used without modifications on any language, as long as a chunker is available for that language.