🎯 Word Bocce

Navigate Semantic Space with Vector Arithmetic

A strategic multiplayer game where you manipulate word embeddings
to reach target words before your opponents

✨ Now live at http://45.79.160.157:8000 ✨

With enhanced visualizations and improved explanations!

Press → or click Next to continue

What Are Word Embeddings?

Every word is represented as a point in 100-dimensional space.

Words with similar meanings are closer together in this space.

📊 Example: "king" and "queen" are close to each other

📊 Example: "king" and "pizza" are far apart

💡 Key Insight: You can do math with these vectors!

This allows for semantic operations like:
king - man + woman ≈ queen

Your Challenge Each Round

You're given a START word and a TARGET word.

Your goal: Navigate from START to TARGET using vector arithmetic.

WATER
→
OCEAN

How do you get there? By adding and subtracting word vectors using the cards you're dealt.

Who wins? The player whose result vector is closest to the target (measured by cosine similarity).

✨ No math skills needed! The computer does all the calculations. You just pick words that feel semantically right.

Your Tools: The Cards

WATER
→
OCEAN
blue
deep
wave
salt
fish
ship
shore
tide
beach
pacific
vast
JOKER

You must pick ONE public and ONE private card to make your move.

How to Make Your Move

WATER
→
OCEAN
  1. 🎯 Choose ONE public card (e.g., "deep")
  2. 🔒 Choose ONE private card (e.g., "vast")
  3. ➕➖ Pick + or − for each card
  4. 🧮 Calculate your result vector:
result = v(water) + v(deep) + v(vast)

Adding a card vector moves you toward that concept.

Subtracting a card vector moves you away from that concept.

Your score: Cosine similarity between your result and "ocean".

Example: Alice's Strategy

WATER
→
OCEAN

Alice chooses:

+ deep
+ vast
v* = v(water) + v(deep) + v(vast)

Her reasoning: "Deep + vast = large body of water = ocean!"

Result: 0.7823 similarity

Nearest word to her result: "seas"

Bob's Competing Strategy

WATER
→
OCEAN

Bob chooses differently:

+ pacific
+ salt
v* = v(water) + v(pacific) + v(salt)

His reasoning: "Pacific is an ocean, salt water = ocean!"

Result: 0.8241 similarity

Nearest word to his result: "oceans" (even better!)

Bob's strategy was more direct → Higher score!

🏆 Round Results

WATER
→
OCEAN
🥇 Bob 0.8241 → "oceans"
🥈 Alice 0.7823 → "seas"

Highest similarity wins!

Bob chose cards that moved him more directly toward "ocean".

What's Really Happening

Words exist as vectors in 100-dimensional space.
Here's a simplified 2D view:

(Actual game uses all 100 dimensions for precise calculations)

Blue: Starting word (water)

Yellow: Target word (ocean)

Purple/Pink: Card vectors you add

Orange: Alice's result (closer to target = higher score)

Strategic Techniques

Master these vector operations to win:

💡 Addition: Combine Concepts

king − man + woman ≈ queen

Add related concepts to navigate semantic relationships

💡 Subtraction: Remove Concepts

france − paris + tokyo ≈ japan

Subtract to move away from unwanted meanings

🃏 JOKER Cards: Ultimate Flexibility

Type ANY word!

Use wildcards to craft the perfect vector path

Why Play Word Bocce?

✨ Learn AI concepts intuitively - Understand word embeddings through gameplay

🧠 Strategic thinking required - Plan your vector path carefully

👥 Outsmart your friends - Compete in real-time multiplayer

⚡ Fast-paced rounds - 2 minutes to make each move

🎯 Easy to learn, hard to master - Simple rules, infinite strategies

🎓 Educational - Used in universities to teach NLP concepts

Two Ways to Play

👥 Multiplayer Mode

Compete against friends in real-time

  • Quick Play: Share link instantly
  • Private cards keep strategy secret
  • 2-minute rounds
  • Live leaderboard

🎯 Puzzle Mode

Solve solo challenges at your own pace

  • 10 curated puzzles
  • Easy, Medium, Hard difficulties
  • Star rating system
  • Hints available

🎯 How Does Scoring Actually Work?

The game uses TWO different methods

📐 Vector Arithmetic

Measures how close you got to the target

Used for: Your score, Best move rankings

result = start + cards
score = similarity(result, target)

🔍 3CosAdd (Analogy Matching)

Finds which word matches your pattern

Used for: "Nearest word" display

score = Σcos(word, positive)
- Σcos(word, negative)

Both are correct! They measure different things.

📐 Method 1: Vector Arithmetic

Think of words as points in space

Step 1: Start at the "start word" position

Step 2: Add your cards like directions:

  • +woman = move toward "woman"
  • -man = move away from "man"
  • +royal = move toward "royal"

Step 3: See where you end up

Step 4: Measure distance to target

final_position = king + woman + royal
score = how_close(final_position, queen)

🎯 This is like navigating: "From the castle, go toward the woman's quarters and toward the throne room"

🔍 Method 2: 3CosAdd (Analogy Matching)

Finds which word fits the pattern

What it does:

Looks at EVERY word in the vocabulary and asks:

"Which word is most similar to [king, woman]
and least similar to [man]?"

How it calculates:

  • Add up similarity to positive words (king, woman)
  • Subtract similarity to negative words (man)
  • Pick the word with highest total score
For each word in vocabulary:
score = similarity(word, king)
+ similarity(word, woman)
- similarity(word, man)

🎯 This is like asking: "What's royal like king, feminine like woman, but not masculine like man?" → Queen!

Example: King → Queen

Let's see both methods in action

KING → QUEEN

📐 Vector Arithmetic

Scores all possible moves:

Move Score Rank
+woman +royal 0.832 #1
+royal +crown 0.818 #3
+woman -man 0.770 #15

✓ +woman +royal gets closest to queen's position in space

🔍 3CosAdd

For move: +woman -man

Finds nearest word:

Word 3CosAdd Score
queen 0.786
monarch 0.699
throne 0.690
princess 0.666

✓ "queen" is the best analogical match!

Why Use Both Methods?

📐 Vector Arithmetic tells you:

"How well did you do?"

Your score of 0.770 means you got 77% of the way to queen.
Someone who played +woman +royal got 83% of the way there.

✓ Fair rankings: Better moves = higher scores

🔍 3CosAdd tells you:

"What concept did you create?"

Your move king+woman-man created the concept of "queen"!
The classic analogy still works perfectly.

✓ Meaningful interpretation: See what you actually made

The Result:

✅ Rankings make sense (closer to target = higher rank)
✅ Classic analogies work (king-man+woman finds "queen")
✅ You get both strategic feedback AND conceptual understanding

Think of it like navigation...

🗺️

Vector Arithmetic

"From the castle, walk 100 steps toward the garden and 50 steps toward the tower."

→ Ends up 200 meters from the fountain

🎯

3CosAdd

"Where you ended up is closest to the library."

→ Names the landmark you reached

In Word Bocce:

📐 Vector math: "You got 77% of the way to queen"
🔍 3CosAdd: "The word you created is 'queen'"

Both pieces of information are useful! 🎯

Ready to Play?

Start navigating semantic space!

Free to play • No download required • Works in any browser

🌐 Hosted at: http://45.79.160.157:8000

1 / 13