Decoding Customer Sentiments: A Dashboard-driven Solution for Product Comment Analysis

Abstract

In the rapidly changing e-commerce landscape, businesses need to use the most advanced analytics technology to gain competitive advantage and meet customer needs. Our search starts with traditional methods such as NLTK + TextBlob and VADER, which provide a stable baseline with an accuracy of 73.99% (F1 66.01%) and 67.35% (F1 54.46%), respectively. However, in search of a deeper understanding, we chose Word2Vec, which uses single word embeddings to capture relationships, achieving 80.01% accuracy (F1 79.56%), making the hypothesis clear. The LSTM model, known for its connectivity pattern recognition, showed 69.95% accuracy (F1 66.35%), but its performance paled in comparison to Transformer-based architectures. Among them, BERT achieved 68.57% (F1 73.93%) accuracy despite the challenges of computing resources and good knowledge training, changing the perception of user feedback and analysis. However, the undisputed winner in determining user sentiment, surpassing all other models with an impressive 97.70% accuracy and 86.21% F1 score, is the BiLSTM model, which refers to the use of technology to create content and gather opinions from users to increase traffic. The importance of providing useful feedback to customers, business success and e-commerce growth.