Lacking of full exploitation of the semantic and sentiment feature of review, which is crucial for revealing the real attention of reviewer, hence we propose a Multi-Semantic Learning and Sentiment-aware (MSLS) model for deceptive review detection. A multi-semantic learning model based on a shared RoBERTa is proposed, which achieves a comprehensive capture of multi-level semantic information of review by extracting the left local semantic features, the right local semantic features, and the global semantic features of reviews parallelly. Furthermore, to obtain the sentiment awareness ability and mitigate of feature conflicts, a two-stage feature fusion structure is designed. In the first fusion stage, the unified sentiment representation generated by pre-training model SKEP is deeply fused with the global semantic feature through BiLSTM with the attention mechanism; in the second stage, the deeply fused feature is concatenated with the left and right local semantic feature. Based on the YelpZip and YelpChi dataset, the comparative experiments with other baseline methods of deceptive review detection are conducted. The experimental results show that MSLS outperforms
other deceptive review detection methods in different scenarios, the recall on YelpZip and YelpChi dataset improved by 4.61% and 4.40%, which demonstrates that multisemantic and sentiment enhancement are helpful for deceptive review detection.