A new LSTM network model combining TextCNN

Abstract

The development of computer communication technology has brought massive amounts of spam texts. Spammers use a variety of textual means to avoid detection of spam texts, which has brought challenges to spam text filtering technology. Deep neural network has superior performance in feature representation and feature extraction. TextCNN based on convolutional neural network can extract the local feature representation of sentences, but ignore the successive relationship between words. The LSTM based on the recurrent neural network takes into account the sequential relationship between words, but it is not as good as TextCNN in representation of local features. We propose an algorithm that combines the TextCNN and LSTM network called TC-LSTM to implement spam text filtering, and compare the Precision, Recall and F-measure indicators with the traditional TextCNN and LSTM on two datasets. Experiments show that our TC-LSTM algorithm is superior to the traditional TextCNN and LSTM networks in spam text filtering.