title	Sentiment Lab
emoji	🔮
colorFrom	pink
colorTo	indigo
sdk	docker
app_port	7860
pinned	false

Sentiment Analysis on Mobile Phone Reviews

A comparative study of five machine learning models for classifying sentiment in mobile phone reviews, built on a custom-scraped corpus of 10,000+ reviews from GSMArena. The project trains, evaluates, and serves predictions through a Flask web dashboard, comparing two independent labeling pipelines (a classical rule-based approach and a transformer-based approach) across the same data.

Live demo: huggingface.co/spaces/UroojFatima-22/Sentiment-Analysis

What this project does

• Scrapes user reviews for popular phones across four brand categories (Samsung, iPhone, Xiaomi, Mixed)
• Cleans the raw text through an 11-step preprocessing pipeline
• Generates two parallel sets of sentiment labels using independent methods:
  • VADER + TextBlob (lexicon-based, fast)
  • RoBERTa transformer (cardiffnlp/twitter-roberta-base-sentiment-latest)
• Trains five different classifiers on each labeled dataset:
  • Multinomial Naive Bayes
  • Logistic Regression
  • Linear SVM
  • K-Means clustering (unsupervised, with majority-vote cluster mapping)
  • Bi-LSTM RNN (PyTorch)
• Evaluates every model on a held-out test set with accuracy, precision, recall, F1, and confusion matrices
• Serves a Flask web app with two interfaces:
  • A dashboard for running individual models or comparing all five
  • A live prediction page that shows what all five models predict for a typed-in review, in real time

Tech stack

• Python 3.10+
• scikit-learn for the classical models, TF-IDF, and K-Means
• PyTorch for the Bi-LSTM RNN
• Transformers (Hugging Face) for the BERT-based labeler
• NLTK, VADER, TextBlob for preprocessing and rule-based labeling
• Flask for the web app
• Chart.js for the comparison visualization

Project structure

.
├── app/
│   ├── app.py                  # Flask backend (routes + model loading)
│   ├── templates/              # index.html + predict.html
│   └── static/css/style.css    # Dashboard styling
├── scrapers/
│   ├── gsmarena_scraper.py     # Scrapes review pages
│   └── phone_urls.py           # Phone URL configuration
├── preprocessing/
│   ├── parser.py               # raw_txt -> parsed_reviews.xlsx
│   ├── cleaner.py              # Text cleaning pipeline
│   ├── labeler.py              # VADER + TextBlob labeler
│   └── labeler_bert.py         # RoBERTa labeler
├── models/
│   ├── train.py                # Trains all 5 models on both datasets
│   └── evaluate.py             # Evaluates everything on the test set
├── data/                       # Generated by the pipeline (gitignored)
│   ├── raw_txt/                # Scraped reviews per brand
│   └── processed/              # Excel files at each pipeline stage
├── config.py                   # Central configuration
└── requirements.txt

Setup

This project does not ship with data or trained models. You'll need to run the pipeline once to generate them.

1. Install dependencies

python -m venv venv
venv\Scripts\activate          # Windows
# source venv/bin/activate     # macOS/Linux

pip install -r requirements.txt

2. Download NLTK resources

python -m nltk.downloader punkt punkt_tab stopwords wordnet

3. Run the pipeline

The full pipeline produces the data and models the app needs. Run these in order:

# 1. Scrape reviews from GSMArena (slow — uses a polite delay between requests)
python scrapers/gsmarena_scraper.py

# 2. Parse the raw .txt files into a structured Excel file
python preprocessing/parser.py

# 3. Clean the text (lowercase, remove URLs, handle contractions, lemmatize, etc.)
python preprocessing/cleaner.py

# 4a. Label with VADER + TextBlob (fast)
python preprocessing/labeler.py

# 4b. Label with RoBERTa (slow — uses GPU if available)
python preprocessing/labeler_bert.py

# 5. Train all 5 models on both labeled datasets
python models/train.py

# 6. (Optional) Evaluate all 10 trained models on their held-out test sets
python models/evaluate.py

After step 5, you'll have ten saved models in total — five per labeling pipeline, in models/saved_vader/ and models/saved_bert/.

4. Run the web app

python app/app.py

Open http://127.0.0.1:5000 in your browser.

Using the app

Dashboard (/)

Pick a dataset, label source (VADER or BERT), brand filter, and a model. Click Run Model to see accuracy, precision, recall, F1, and the confusion matrix on the held-out test set. Click Compare All 5 to render a side-by-side bar chart of every model's metrics.

All metrics are computed on the 20% test split that the models never saw during training, so the numbers reflect actual generalization rather than memorization.

Live Predict (/predict)

Type any review. The app runs all five models on your input and shows a consensus view: which sentiment the majority predicted, how the votes split, and each individual model's verdict. A cleaned-text preview shows what the models actually received after preprocessing.

Configuration

All paths, hyperparameters, and thresholds are centralized in config.py. Notable settings:

• TEST_SIZE = 0.2 — train/test split ratio
• RANDOM_STATE = 42 — deterministic splitting across all scripts
• POSITIVE_THRESHOLD / NEGATIVE_THRESHOLD — VADER compound score cutoffs for label assignment
• MODELS_DIR_VADER / MODELS_DIR_BERT — output directories for each labeling pipeline

Notes

• The pipeline is fully reproducible: same random_state everywhere, same TF-IDF settings shared between classical models and K-Means, same hyperparameters in train.py.
• BERT labels are higher-quality (manual audit indicates ~83% accuracy on this corpus) but harder to predict from bag-of-words features, so BERT-trained models show lower raw accuracy than VADER-trained ones. Macro-F1 is the more honest metric for comparing across labeling pipelines because it weights all three sentiment classes equally rather than rewarding majority-class predictions.
• The Flask app loads all ten models at startup, which takes a few seconds. After that, predictions and evaluations are fast.

---

License

MIT © Urooj Fatima

Author

Built by Urooj Fatima — LinkedIn · Email