Word-Embedding Model for Evaluating Text Generation of Imbalanced Spam Reviews

Generating minor class data of spam texts is expected to solve the imbalanced problem in spam detection of product reviews. There could be semantic differences between the generated texts and the original ones. Thus, by including the semantically differed texts in the spam dataset used for training is like a noise addition. For evaluating the generated texts, some manual preparations of ground-truth data are necessary. This work has evaluated the generated texts with some approaches to ensure their context and sequence similarities compared to the original texts for better performance of a spam detection. The employed approaches are expected to eliminate the manual tasks. Our research proposes an evaluation model that consists of word-embedding pre-trained and LSTM Siamese to evaluate text generation in imbalance review. The use of a combination of pre-trained word embedding and LSTM Siamese trained model will capture the semantic aspect of the text.