In this paper, a new topic identification method, WSIM, is investigated. It exploits the similarity between words and topics. This measure is a function of the similarity between words, based on the mutual information. The performance of WSIM is compared to the cache model and to the well-known SVM classifier. Their behavior is also studied in terms of recall and precision, according to the training size. Performance of WSIM reaches 82.4 % correct topic identification. It outperforms SVM (76.2%) and has a comparable performance with the cache model (82.0\%).