Paraphrase Identification Based on ResNeXt

The misuse of paraphrase generation technology has flooded the Internet with low-quality, unclear manuscripts, deteriorating the user’s reading experience. This paper mainly investigates into the paraphrase identification from the deep learning perspective, and proposes a paraphrase identification model based on deep ResNeXt. Our model captures the semantic and syntactic features of the given text with the help of deep neural network. Specifically, we select multiple convolution kernels with different sizes to extract the features of word sequences of different lengths. At the same time, the residual block which helps build a paraphrase identification model suitable for various current platforms. The data preprocessing includes data cleaning, word segmentation, and word vectorization. Experiments on benchmark dataset MSRPC show that our model achieved a strong result of 85.4% on accuracy rate.