An integrated approach: fake review detection using convBERT-BiLSTM classification

Abstract

In the era of E-commerce, online reviews significantly shape consumer buying decisions and store evaluations. However, the prevalence of unethical practices such as review manipulation poses a considerable challenge. Businesses often hire spam reviewers or deploy bots to boost their reputation or even damage that of their competitors. Despite existing efforts in the field of fake review detection, there remains a need for further studies. In contribution, we propose the development of a scoring rubric designed to guide annotators in the identification of fake reviews and a hybrid model ConvBERT-BiLSTM for detection. We leverage the efficiency of ConvBERT, a compact variant of the BERT model, and the superior capabilities of BiLSTM over LSTM. The model is trained on a dataset gathered from Amazon. The dataset comprises 7,727 labeled reviews using the rubric. Through careful assessment, the proposed model garnered an accuracy of 97% surpassing previously established BERT variants.