Neural Networks Merging Semantic and Non-semantic Features for Opinion Spam Detection

In recent years, abundant online reviews on products and services have been generated by individuals. Since customers may refer to relevant online reviews when shopping, the existence of fake reviews can affect potential consumption. Opinion spam detection has attracted widespread attention from both the business and research communities. In this paper, a neural network model combining the semantic and non-semantic features based on the detailed feature exploration is established to detect opinion spams. First, the model learns discourse feature representation with hierarchical attention neural networks which can capture local and global semantic information. And then we synthesis the non-semantic features with multi-kernel convolution neural networks. Finally, the last state vectors of the two-feature learning networks are concatenated and taken as input to the softmax layer for classification. Experiments show that the proposed model is very effective and we get 0.853 AUC which outperforms the baseline methods. Besides, the experiment results on an additional dataset also indicate robustness of this identification model.