🧠 NLP Business Case: Automated Customer Reviews

📌 Overview

This project builds a smart product review analysis and recommendation system powered by Natural Language Processing (NLP) and Generative AI. It automates the end-to-end process of understanding customer sentiment, clustering products, and summarizing key insights—helping businesses make data-driven decisions and enhance product visibility.

---

🚀 Key Features

• ✅ Sentiment Classification: Categorize reviews into Positive, Neutral, or Negative.
• 📦 Product Clustering: Group products into 4–6 meaningful categories using unsupervised learning.
• 🧠 Generative AI Summarization: Summarize reviews and generate product highlights using Large Language Models (LLMs).
• 📇 Product Card Generation: Create structured product summaries that can be integrated into a user interface.

---

🧩 Project Steps

🔹 Part 1: Data Preprocessing & Sentiment Label Classification

🎯 Objective

Clean and prepare the Amazon review data, then classify each review into Positive, Neutral, or Negative categories to gain valuable insights into customer sentiment.

📚 Dataset

We used Amazon product review data from Kaggle:

• Datafiniti_Amazon_Consumer_Reviews_of_Amazon_Products.csv
• Datafiniti_Amazon_Consumer_Reviews_of_Amazon_Products_May19.csv

These were merged to create a more diverse and robust dataset.

⚙️ Workflow Summary

📍 Notebook: data_exploration_and_acquisition.ipynb, data_preprocessing.ipynb

• Load and explore both datasets
• Combine and clean the review data
• Apply sentiment labeling using rule-based or lexicon-based methods (e.g., VADER)
• Save final dataset as cleaned_amazon_reviews_final.csv

---

🔹 Part 2: Sentiment Classification with DistilBERT

🎯 Objective

Use a pretrained transformer model (DistilBERT) to classify Amazon reviews into sentiment categories more accurately than traditional models.

📂 Dataset

• cleaned_amazon_reviews_final.csv with:

  • full_review (cleaned review text)
  • sentiment (labeled sentiment)

⚙️ Workflow Summary

📍 Notebook: Classification_Model_1.ipynb

• Load and encode sentiment labels
• Tokenize using Hugging Face’s DistilBERT tokenizer
• Fine-tune DistilBertForSequenceClassification on labeled reviews
• Evaluate model using accuracy and confusion matrix

✅ Achieved 95.29% accuracy

---

🔹 Part 3: Product Clustering & GPT Summarization

🎯 Objective

Organize products into clusters and generate summaries and recommendations using LLMs like GPT.

🛠️ Workflow Overview

1. Clustering

   • Dimensionality reduction with UMAP
   • Clustering with KMeans (11 clusters)
   • Relabled the Clusters in 4
   • Assign descriptive cluster_labels using GPT-based interpretation

2. Product-Level Aggregation

   • Group reviews by name and brand
   • Compute average rating, sentiment trends, and gather review samples

3. GPT-Based Summary Generation

   • Identify top and worst products per category
   • Output structured summaries in JSON:

4. Product Card Generation

   • Final outputs rendered as Product Cards with:

     • Product name, brand, rating
     • Review highlights and representative image
     • Category and source URL

---

🛠️ Tech Stack

• Python (pandas, numpy, sklearn, matplotlib, seaborn)
• Hugging Face Transformers (DistilBERT)
• UMAP, KMeans
• OpenAI GPT-4 API (for summarization)
• Jupyter Notebooks

---

📈 Results

• ✅ 95.29% accuracy on sentiment classification using DistilBERT.
• 🔍 Products grouped into coherent clusters such as Kindle, Fire Tablets, Accessories.
• 💬 GPT-generated summaries provide actionable insights and highlight standout products.