Reinforcement of Pre-trained Bert Architecture for the Detection of Spam Reviews

Abstract:
Spam reviews misguide consumers' decisions and may seriously influence transactions in the online markets. Existing detection modules mainly rely on tedious by-hand characteristics engineering, which expects much professional knowledge. Recent efforts utilize deep learning to extract semantics knowledge, and these methods cannot consider comprehensive potential semantics well. We propose a novel, reinforced pre-trained Bert architecture for spam review detection, including a multilayer bidirectional Trans-former encoder (BERT) and a 3-Layer fully connected network (3FCN) to capture comprehensive, essential, and meaningful information. Specifically, the BERT learns comprehensive semantic information, and the 3FCN reinforces the pre-trained Bert to obtain essential and meaningful information. Experiment results show that the proposed architecture has better detection performance.