Advanced Product Helpfulness Detection Using BERT and LSTM

In today’s global marketplace, consumers are exposed to a vast volume of product reviews across numerous platforms, making it increasingly difficult to identify genuinely helpful feedback. Reliable review insights are crucial not only for consumers making purchase decisions but also for businesses aiming to improve product quality, service, and overall customer satisfaction. This research introduces a machine learning framework designed to evaluate the helpfulness of product reviews by leveraging advanced natural language processing techniques. Utilizing the Amazon Fine Food Reviews dataset, we propose a novel feature engineering approach named BERF (BERT with Random Forest probabilities), which combines contextual BERT embeddings with class probability outputs. The dataset imbalance is addressed using the Synthetic Minority Oversampling Technique (SMOTE), and multiple classification algorithms are employed. Among them, the Light Gradient Boosting Machine (LGBM) achieved the highest accuracy of 98%, outperforming existing models. The model performance is validated through k-fold cross-validation and optimized using hyperparameter tuning. Our findings demonstrate that the proposed approach effectively enhances the prediction of review helpfulness, potentially minimizing misinformation and supporting more informed online purchasing decisions.