Multi-Label Intent Classifier for Google Play Store App Reviews

A Natural Language Processing (NLP) project submitted for CS6320 - Natural Language Processing (Spring 2025) at The University of Texas at Dallas.

By: Rishika Vaish
Instructor: Dr. Tatiana Erekhinskaya

---

Project Demo

Watch the 5-minute project walkthrough and demo on YouTube:
👉 Click to Watch

---

Project Overview

This project analyzes user reviews from Google Play Store apps and automatically classifies them by:

• Sentiment (Positive, Neutral, Negative)
• Multiple Intents per review (e.g., Recharge issue + Customer care problem)

The goal is to help app developers extract meaningful, categorized insights from unstructured textual feedback, enabling quicker resolution and product improvements.

---

Key Features

• Preprocessing using NLTK (tokenization, stopword removal, lemmatization)
• Sentiment Analysis via VADER
• Clustering with TF-IDF + K-Means to pre-label reviews
• Multi-Label Classification using fine-tuned XLNet model
• Real-Time Web App built with Streamlit
• Confidence Scores are shown for each detected label

---

Intent Labels Detected

The model identifies one or more of the following intents:

• Problem with recharge
• Problem with reward/redeem points
• Problem in customer care service
• Problem in registration/login
• Other complaints
• Appreciation
• Bad/Irrelevant comments

---

Sample Screenshots

Web Interface

Classification Result Example

---

Getting Started

Installation

Clone the repo:

git clone https://github.com/rishika7006/intent-classifier-xlnet.git
cd intent-classifier-xlnet

Install dependencies:

pip install -r requirements.txt

Run the Streamlit App

streamlit run 03_multilabel_classifier_app.py

Related repo: sentiment-intent-app — Streamlit dashboard that applies this XLNet classifier end-to-end on scraped Play Store reviews.

---

Testing & Feedback

Testing was conducted with 5 independent users.
🔹 Positive feedback on ease of use and multi-intent detection
🔹 Suggested improvements:

• Add progress tracker
• Speed up backend
• More detailed intent descriptions

---

Lessons Learned

• Balancing model performance with real-time speed is crucial for web apps
• XLNet worked effectively for capturing nuanced, multi-intent review content
• TF-IDF clustering gave a strong starting point for label design
• User feedback was essential for refining app usability

---

Final Report

📄 You can find the full project report here

---

Contact

Feel free to reach out if you'd like to collaborate or discuss the project!

Rishika Vaish
CS6320 - Natural Language Processing
The University of Texas at Dallas