Optimizing semantic LSTM for spam detection

Abstract

Classifying spam is a topic of ongoing research in the area of natural language processing, especially with the increase in the usage of the Internet for social networking. This has given rise to the increase in spam activity by the spammers who try to take commercial or non-commercial advantage by sending the spam messages. In this paper, we have implemented an evolving area of technique known as deep learning technique. A special architecture known as Long Short Term Memory (LSTM), a variant of the Recursive Neural Network (RNN) is used for spam classification. It has an ability to learn abstract features unlike traditional classifiers, where the features are hand-crafted. Before using the LSTM for classification task, the text is converted into semantic word vectors with the help of word2vec, WordNet and ConceptNet. The classification results are compared with the benchmark classifiers like SVM, Naive Bayes, ANN, k-NN and Random Forest. Two corpuses are used for comparison of results: SMS Spam Collection dataset and Twitter dataset. The results are evaluated using metrics like Accuracy and F measure. The evaluation of the results shows that LSTM is able to outperform traditional machine learning methods for detection of spam with a considerable margin.