This paper introduces the method of collocation of confusion words and prepositions based on Winnow algorithm. Firstly, we obtain training sets by confusion sets. After preprocessing the training sets, we use the text feature extracted method to obtain feature sets. Secondly, we get the feature sets with weights by training on the feature sets based on Winnow and extract the prepositions which appear after the confusion words to generate the preposition vectors. Finally, we extract features from the test sets and the test sets and get the real-word errors checking results by the test which combines Winnow algorithm and the method of collocation of confusion words and prepositions. The correct rate, recall rate and F1 measure of some confusion words are improved by 10% ~ 20% when we join the method of collocation of confusion words and prepositions, some even up to 100%.