Simple Language Models for Spam Detection

For this year’s Spam track we used classifiers based on language models. These models are used to compute the log-likelihood for each individual message and then classify them as either ham or spam. Different data sets were used to train these language models. Our approach is simple, we initially create simple unigram language models and smooth the probabilities of unseen tokens by means of the expected likelihood estimator with a small discount probability tuned in a training corpus.