This project performs intelligent classification of customer feedback using Natural Language Processing (NLP). It classifies over 8,500 manually labeled reviews into multiple categories such as sentiment, recommendation, and query using DistilBERT, a transformer-based model.
We developed a system that processes real-world feedback and performs the following tasks:
Sentiment Analysis – Identifies whether the feedback is positive, negative, or neutral.
Recommendation Detection – Flags feedback containing suggestions for improvement.
Query Detection – Detects whether the message is a user query instead of regular feedback.
The model is trained using a manually curated dataset of 8,500+ labeled reviews, ensuring domain relevance and annotation quality.
Our model performs multi-label classification on each review to identify the following:
✅ Sentiment Classification Determines whether the feedback expresses a Positive, Negative, or Neutral sentiment. Example: ➤ “The service was excellent and the response was quick.” → Sentiment: Positive
💬 Recommendation Detection Identifies if the feedback includes a suggestion or feature request from the user. Example: ➤ “You should include a PDF export option.” → Recommendation: Yes
❓ Query Detection Detects whether the feedback is a question, help request, or not an actual review. Example: ➤ “How do I change my password?” → Query: Yes
Each review may trigger one or more labels simultaneously, allowing richer and more contextual understanding.
Language: Python
NLP Model: DistilBERT from HuggingFace Transformers
Libraries:
transformers
pandas, numpy, scikit-learn
streamlit – for the interactive dashboard
oracledb – to connect and fetch reviews from an Oracle SQL database
8,500+ manually labeled customer reviews
Format: CSV (new_corrected_reviews.csv)
Sahil Sohani Vaibhavi Bhardwaj