Unsupervised Approaches to Text Correction Using Google N-grams for English and Romanian

We present an unsupervised approach that can be applied to test corrections tasks such as real-word error correction, near-synonym choice, and preposition choice, using n-grams from the Google Web 1T dataset. We present in details the method for correcting preposition errors, which has two phases. We categorize the n-gram types based on the position of the gap that needs to be replaced with a preposition. We also consider the normalized frequency values of the candidate prepositions. Our experimental results are better than those reported in related work, on the same test set that was also used in related work. We applied our method to English, but it can be easily applied to Romanian. Google released n-gram counts for 10 European languages, including Romanian. As test set, a part of a Romanian corpus can be used. A subset of prepositions needs to be chosen. If we include prepositions that consist in two words, the algorithm needs to be adjusted.