MIDS Capstone Project Summer 2023

Review Sentinel

Unlock the Power of Negative Customer Reviews for Unprecedented Growth!

ReviewSentinel specializes in extracting invaluable insights from negative Amazon customer reviews to spark transformative business growth. Our advanced NLP techniques uncover hidden potential within dissatisfaction, empowering sellers to refine offerings and achieve a competitive edge. For existing sellers, our comprehensive feedback analysis illuminates areas for improvement, enabling data-driven decisions that elevate customer satisfaction and drive remarkable expansion. Moreover, we empower potential buyers with nuanced insights from negative reviews, guiding confident purchasing decisions and unparalleled shopping experiences.

We used the Amazon Review Data ranging from May 1996 through Oct 2018 and selected four distinct products (Refrigerator, Ladder, Standing Desk, and Lawn Mower) under 4 product categories (Appliances, Home & Kitchen, Office Products, Patio Lawn & Garden) respectively for our project. Our core task focuses on text classification to categorize negative Amazon reviews. However, as our dataset was inherently unlabelled, we developed our own labelling methodology and manually labeled our dataset into themes such as Quality, Design/Functionality, Delivery/Packaging and Other. We explored models such as Naive Bayes and fine tuned LLM models which includes BERT base uncase and BERT large uncase and compared the model options primarily based on F1-score. To further enhance the insights on delivery, we ventured into topic modeling leveraging unsupervised models like BERTopic, prioritizing multi-modal approaches for relevance. Our approach strives to minimize error impact and streamline interpretation for sellers. Looking ahead, we would like to explore text classification's generalizability and refine text generator models for improved topic clustering interpretation.

Last updated: August 8, 2023