Semantic Indexing-Based Data Augmentation for Filtering Undesired Short Text Messages

Abstract

In the last years, spammers have taken advantage of the popularity of electronic media to spread undesired text messages. These may cause direct and indirect damages, such as dissatisfaction and exposure of users to misleading information and malicious content that can result in significant financial losses. Automatic filtering short text messages is a challenging problem nowadays because labeled datasets generally contain few instances and messages may have an insufficient amount of terms to be classified. In addition, the messages are rife with abbreviations, slang, and misspelled words making it difficult to generate a good computational representation. In this study, we propose an automatic data augmentation technique to increase the number of labeled instances and to improve the quality of the computational representation of short and noise text messages. We also proposed an ensemble approach to combine the predictions obtained by the classifiers using the messages generated by this technique. Experiments with three text representation techniques demonstrated that the ensemble approach improves the results obtained in the detection of undesired short text messages when the number of training instances is small.