Effectiveness of Fine-tuned BERT Model in Classification of Helpful and Unhelpful Online Customer Reviews

The problem of information overload in online review platforms has seriously hampered many customers’ ability to evaluate the quality of products or businesses when making purchasing decisions. A large body of literature exists that attempts to predict the helpfulness of online customer reviews and has reported contradictory findings on the effectiveness of various approaches. Moreover, many existing solutions use traditional machine learning techniques and handcrafted features, limiting generalization. Therefore, this study aims to propose a generalized approach by fine-tuning the BERT (Bidirectional Encoder Representations from Transformers) base model. The performance of BERT-based classifiers is then compared with that of bag-of-words approaches to determine the effectiveness of BERT-based classifiers. The evaluations performed using Yelp shopping reviews show that fine-tuned BERT-based classifiers outperform bag-of-words approaches in classifying helpful and unhelpful reviews. In addition, it is found that the sequence length of the BERT-based classifier has a significant impact on classification performance.