Insurance Review Rating Prediction

📌 Project Overview

This project aims to analyze customer reviews of insurance services using Natural Language Processing (NLP) techniques. The primary goal is to predict customer ratings based on textual feedback and extract key insights to improve insurance services. A web application is also deployed using Streamlit to allow users to input their reviews and receive predicted ratings.

🎯 Objectives

• Classify customer reviews into predefined categories or themes.
• Predict ratings from textual feedback.
• Extract insights from customer opinions to enhance service quality.
• Deploy an interactive web app for real-time predictions.

🗂 Dataset

• Source: 35 text files (~300 KB each) containing customer reviews.
• Features:
  • DateTime: Time of the review.
  • Insurer: Name of the insurance provider.
  • Type: train or test dataset.
  • Review: Customer feedback in French (translated to English).
  • Rating: Numeric rating associated with the review.

🛠️ Methodology

1️⃣ Data Preprocessing & Cleaning

• Removing URLs, HTML tags, emojis, and special characters.
• Expanding contractions (e.g., can't → cannot).
• Correcting spelling errors.
• Tokenization, stopword removal, and lemmatization.

2️⃣ Feature Engineering

• N-gram Analysis: Extracting frequent word patterns.
• Topic Modeling: Using LDA to group reviews into themes.
• Word Embeddings: Using Word2Vec (W2V) and clustering with k-means.

3️⃣ Modeling

• Baseline: TF-IDF + Logistic Regression.
• Neural Network: Basic NN with an embedding layer.
• Fine-Tuned Transformer Models:
  • RoBERTa (GPT-2)
  • LLaMA 3.2 (1B parameters) with LoRA fine-tuning
• Final Choice: Basic NN (small size, high accuracy, easy to deploy).

4️⃣ Model Evaluation

• Metrics: Accuracy, Precision, Recall, F1-score.
• SHAP Analysis: Explainability of model predictions.

5️⃣ Deployment

• Streamlit Web App for real-time review analysis.
• Hosted at: Insurance Opinion Classification App

📊 Results & Insights

• Key Influencing Factors:

  • Terms like "speed", "premium", and "high price" significantly impact ratings.
  • Negative reviews are more focused on pricing and customer service delays.
  • Positive reviews mention fast claims processing and good customer support.

• Model Performance:

  • Selected Model: Basic NN (Embedding + 1D Convolution)
  • Accuracy: ~85%
  • Mean Distance from True Rating: ~0.75 (good predictive capability)

📌 Future Improvements

• Integrating French-specific embeddings for better semantic understanding.
• Implementing attention-based mechanisms for more refined predictions.
• Expanding the dataset with more diverse customer feedback sources.

🏆 Contributors

• Ahmed Maaloul
• Martin Pujol