McDonald's Sentiment Analysis

A lightweight NLP pipeline that mines ~33k public Google reviews to help stakeholders quickly gauge customer sentiment across U.S. locations.

• Manual review reading is slow and inconsistent.
• Automate sentiment scoring and aspect tagging for food, service, speed, and cleanliness.
• Provide state-level insights for decision-makers.

• Focus on ~33k English Google reviews from U.S. stores.
• Class project delivered over one semester.
• Aspect keywords curated manually; no PII retained.

• Python ETL with Pandas for cleaning and aggregation.
• NLTK + VADER for sentiment scoring.
• Aspect tagging via keyword rules.
• Charts prototyped in Orange Data Mining.

• Source: 33k Google reviews with location metadata.
• Cleaning: deduplication, lowercasing, stop-word removal, and tokenization.
• Sentiment thresholds: positive ≥ 0.05, negative ≤ -0.05, else neutral.

• Bulk review ingestion and preprocessing.
• VADER sentiment scoring per review.
• Aspect classification for food, service, speed, and cleanliness.
• State-level summaries and charts.

• VADER polarity scores binned into positive, neutral, negative.
• Aspect frequency counts to spot recurring themes.
• Manual spot checks for validation.

Automated processing replaced weeks of manual review reading and highlighted states where service sentiment lagged, informing future improvements.

• Noisy, emoji-filled text required custom cleaning.
• Aspect keywords needed tuning to avoid false positives and negatives.

• Expose results via interactive dashboard or API.
• Experiment with transformer models for aspect sentiment.

• Project Documentation (PDF)