People often use the internet in order to express their opinions for specific issues or to get some information. Flames among online messages disrupt those uses. In this paper, I propose a heuristic method which detects flames from online messages automatically using an n-gram language model. We focus on flaming in Korean web sites, but our system can be applied to any other languages. I propose a method to extract features based on n-grams and score each feature by a heuristic method. The proposed algorithm outperforms a wordbased algorithm in terms of the accuracy and the recall rates, because the algorithm presented in this paper can solve the two problems: variants of words and abbreviations of blanks. In the evaluation, I compare the proposed method with the word-based algorithm and the algorithm based on an n-gram language model which use SVM learning machine. While the proposed algorithm does not need any stemming and tagging tasks, it can detect more accurately by 10% than the algorithm based on words.