Deceptive reviews of products influence customer’s decision and damage the reputation of the product. Most of the existing methods for detecting fraudulent comments use feature design, which is difficult to learn the potential semantics of comments. We propose a neural network based on clustering and attention mechanism to learn the semantic representation of reviews. Specifically, we use DBSCAN to discover the semantic groups in the word embedding space, and then construct the semantics of different semantic groups through the attention mechanism. The model computes the representations of the semantic units and combine them into the sentence representation. In feature selection, we perform feature combination to improve performance. Then we demonstrate the effectiveness of the proposed model through Amazon dataset experiments. In the experiment, the model surpassed the state-of-art method.