Research and implementation of false review recognition based on semantic and non-semantic fusion of migrated sentiment features

Consumers tend to shop online in the era of e-commerce, while some unscrupulous merchants intentionally post false reviews to influence consumer decision-making and undermine fair competition in the market. Therefore, it is crucial to accurately identify false reviews. In this paper, we propose the RBLADC (RoBERTa_BiLSTM_Attention_DANN_CNN) false review recognition model, and the main contributions include: firstly, we propose the LightGBM-RF(Light Gradient Boosting Machine with Random Forest) non-semantic feature selection method, which extracts the set of non-semantic features from the linguistic and non-linguistic information features, and through which we obtain the importance ranking, select the optimal subset, reduce the noise interference, and improve the performance of the model; secondly, we construct the RoBERTa-BiLSTM-Attention semantic feature extraction model, using RoBERTa (Robustly Optimized BERT Pretraining Approach) as the word embedding layer for semantic vector representation, BiLSTM(Bidirectional Long ShortTerm Memory) to learn the deep semantics of the context, and the attention mechanism to capture the important information and improve the semantic feature extraction ability; lastly, using DANN (Domain-Adversarial Neural Network) to extract migrated emotional features, given the emotional tendency of false comments, the comment text is divided into three parts as the target domain, learning shared emotional knowledge to extract fine-grained emotion information and improve the recognition effect. In this paper, semantic, non-semantic, and sentiment feature vectors are spliced to act on the RBLADC model, and experiments show that its accuracy reaches 92.38%, with excellent classification performance.