Language Games

Dead simple semantic word games powered by word vectors, played in the terminal.

Uses GloVe vectors from Stanford NLP with a lightweight pure-numpy backend — no heavy ML frameworks needed. Models download automatically on first run and are cached for instant reloading.

Games

#	Game	How it works
1	Competitive Guessing	A secret word is chosen. Semantic hints are shown. Guess the word — cosine similarity is your score.
2	Closest Word	Pick the word from a list that is most similar to a target.
3	Odd One Out	Find the word that doesn't belong in a group.
4	Semantic Scrabble	Form words from random letter tiles that are semantically close to a target.
5	Analogies	A is to B as C is to ??? — classic word-vector arithmetic.
6	Word Chain	Build a chain of related words. Break the chain and lose a life.
7	Hot & Cold	Guess the secret word guided by a temperature meter — hotter means closer.
8	Category Sprint	Name as many words as you can that belong to a category.

All games support any number of players and three difficulty levels (easy / medium / hard). A session leaderboard tracks wins across games.

Install

uv venv .venv && source .venv/bin/activate && uv pip install -r requirements.txt

Or with pip:

python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt

Run

python play_game.py

On first launch you choose a vector dimensionality (50d–300d). The GloVe zip (~822 MB) downloads once with a progress bar and is cached in ~/.cache/language-games/.

Features

• Animated TUI — similarity bars fill up in real-time, countdowns before reveals, winner celebrations
• Temperature meter — the Hot & Cold game shows a thermometer from Frozen to ON FIRE!
• Letter tiles — Semantic Scrabble renders Scrabble-style tiles
• Session leaderboard — tracks wins across games in a single session
• Difficulty levels — affects similarity thresholds, number of hints, and guess limits
• Topic filtering — optionally restrict the word pool to a semantic topic
• Zero compilation — pure numpy backend, works on any Python 3.10+

Screenshots

Title & Model Selection

Game Menu

Game 1 — Competitive Guessing

Game 2 — Closest Word

Game 3 — Odd One Out

Game 4 — Semantic Scrabble

Game 5 — Analogies

Game 6 — Word Chain

Game 7 — Hot & Cold

Game 8 — Category Sprint

Final Scores & Session Leaderboard

How It Works

Words are represented as dense vectors in a high-dimensional space (50–300 dimensions) where semantically similar words are closer together.

The WordVectors class is a lightweight, pure-numpy implementation that supports:

• Cosine similarity between any two words
• Most-similar lookup using fast argpartition
• Outlier detection by distance from group centroid
• Vector arithmetic for analogies (king - man + woman = queen)

Dependencies

Just numpy and rich.

License

MIT