🎮 Game Overview
You're a data scientist at Amazon tasked with building the ultimate review sentiment classifier. Your challenge: correctly identify positive, negative, and edge-case reviews to train an AI that will process millions of customer opinions. But watch out—sarcasm, mixed feelings, and clever wordplay will test your detective skills!
📚 Phase Details
Phase 1: Human Training
You receive 20 product reviews without labels. Your task:
- Classify each as Positive, Negative, or Neutral
- Flag potential sarcasm or mixed sentiment
- Note key sentiment indicators (words/phrases)
- Build intuition for patterns
Time Pressure: 30 seconds per review!
Phase 2: Algorithm Design
Choose your text analytics approach:
- Bag of Words: Count positive/negative terms
- TF-IDF: Weight important words
- N-grams: Capture phrases like "not bad"
- Sentiment Lexicons: Use pre-built dictionaries
Build rules based on your Phase 1 observations!
Phase 3: Prediction Challenge
Your model faces 80 unseen reviews:
- Apply your algorithm to new data
- No manual intervention allowed
- Reviews include various products
- Hidden test set has tricky cases
Goal: Beat 75% accuracy baseline!
Phase 4: Edge Case Bonus
Special challenge reviews worth double points:
- Sarcastic reviews ("Great, another broken product")
- Mixed sentiment ("Love the design, hate the price")
- Negations ("Not disappointed at all!")
- Context-dependent ("Sick headphones" - good or bad?)
Correctly identify these for bonus points!
🎭 Sample Reviews You'll Encounter
🛠️ Available Techniques
Word Lists Easy
Method: Count positive/negative words
Example: "amazing" (+1), "terrible" (-1)
Accuracy: ~65-70%
Pitfall: Misses context and negations
TF-IDF Medium
Method: Weight words by importance
Example: "broken" matters more than "the"
Accuracy: ~75-80%
Pitfall: Still misses word relationships
N-grams Medium
Method: Capture word pairs/triplets
Example: "not bad" → positive
Accuracy: ~78-83%
Pitfall: Exponential feature growth
VADER Easy
Method: Pre-built sentiment analyzer
Example: Handles emphasis (!!!) and emojis
Accuracy: ~80-85%
Pitfall: Generic, not product-specific
Machine Learning Hard
Method: Train classifier on features
Example: Naive Bayes, SVM, Neural Net
Accuracy: ~85-92%
Pitfall: Needs lots of labeled data
Hybrid Approach Hard
Method: Combine multiple techniques
Example: VADER + custom rules + ML
Accuracy: ~88-95%
Pitfall: Complex to implement quickly
📊 Scoring System
Detailed Scoring Breakdown
- Base Accuracy (500 pts): 5 points per correctly classified review
- Speed Bonus (200 pts): First team to submit gets full points, -20 per rank
- Sarcasm Detection (100 pts): 20 points per correctly identified sarcastic review
- Mixed Sentiment (100 pts): 25 points per correctly handled mixed review
- Innovation Bonus (100 pts): Creative approaches, good documentation, clean code
🏆 Accuracy Thresholds
- < 60%: "Needs Coffee" - Random guessing territory
- 60-70%: "Junior Detective" - Better than random!
- 70-80%: "Senior Analyst" - Solid performance
- 80-90%: "Sentiment Master" - Professional level
- > 90%: "Algorithm Whisperer" - Are you cheating? 😉
⚡ Tricky Cases to Watch For
🎭 Sarcasm Signals
- "Great" followed by negative context
- Excessive punctuation (!!!???)
- Quotes around 'positive' words
- "Thanks for nothing" patterns
🔄 Negation Patterns
- "Not bad" → Actually positive
- "Can't complain" → Positive
- "Wasn't disappointed" → Positive
- "No issues whatsoever" → Very positive
🎯 Domain-Specific Terms
- "Sick" (headphones) → Positive
- "Killer" (app) → Positive
- "Addictive" (game) → Positive
- "Cheap" → Context-dependent!
⚖️ Mixed Signals
- Product good, service bad
- Love X, hate Y patterns
- Star rating vs. text mismatch
- Comparison reviews (better than X, worse than Y)
📊 Rating Inconsistencies
- 5 stars but complaints in text
- 1 star but "not that bad"
- 3 stars could be anything!
- Cultural differences in rating
🌍 Cultural Context
- British understatement
- American enthusiasm
- Technical jargon variations
- Generation-specific slang
🏆 Leaderboard Categories
| Award | Description | Prize |
|---|---|---|
| 1 Accuracy Champion | Highest overall classification accuracy | 500 bonus XP |
| 2 Sarcasm Detector | Best at identifying sarcastic reviews | 300 bonus XP |
| 3 Speed Demon | First to submit with >70% accuracy | 200 bonus XP |
| 🎨 Creative Coder | Most innovative approach | Special recognition |
| 📈 Most Improved | Biggest gain from manual to automated | Encouragement award |
💡 Pro Strategies
🎯 Quick Wins
- Start with simple word counting
- Handle "not" + positive word cases
- Use star ratings as a hint (but don't trust completely)
- Look for ALL CAPS for emphasis
- Check first and last sentences (usually summary)
🚀 Advanced Tactics
- Build product-specific lexicons
- Weight recent reviews more heavily
- Detect review bombing patterns
- Use emoji sentiment (😍 vs 😡)
- Ensemble multiple approaches
⚠️ Common Pitfalls
- Over-relying on single words
- Ignoring context around sentiment words
- Missing double negatives
- Treating all products the same
- Forgetting about neutral reviews
🎓 Learning Objectives
By Playing This Game, You'll Learn:
- Text Preprocessing: Why cleaning and tokenization matter
- Feature Engineering: Converting text to numbers for analysis
- Sentiment Complexity: Why sentiment isn't just positive/negative
- Context Importance: How word relationships change meaning
- Model Limitations: When automated systems fail and need human help
- Real-World Challenges: Sarcasm, irony, and cultural differences
- Performance Metrics: Accuracy, precision, recall, and F1 scores
- Business Impact: How sentiment analysis drives product decisions
🚀 Ready to Start Detecting?
Choose your implementation and become a Sentiment Detective!