Abstract
A ready set of commonly confused words plays an important role in spelling error detection and correction in texts. In this paper, we present a system named ACE (Automatic Confusion words Extraction), which takes a Chi-nese word as input (e.g., " 不脛而走 ") and automatically outputs its easily confused words (e.g., " 不徑而走 " , " 不逕而走 "). The purpose of ACE is similar to web-based set expansion – the problem of finding all instances (e.g. " Halloween " , " Thanksgiving Day " , " Independence Day " , etc.) of a set given a small number of class names (e.g. " holidays "). Unlike set expansion, our system is used to produce commonly confused words of a given Chinese word. In brief, we use some hand-coded patterns to find a set of sentence fragments from search engine, and then assign an array of tags to each character in each sentence fragment. Finally, these tagged fragments are served as inputs to a pre-learned conditional random fields (CRFs) model. We present experiment results on 3,211 test cases, showing that our system can achieve 95.2% precision rate while maintaining 91.2% recall rate.