Play with AI

Inspiration

Word embeddings are a clever way of representing words as vectors. We wanted to get a deeper understanding by building a small word guessing game.

What it does

You and a bot start by imagining a word and the goal for both of you is to arrive at the same word. You consecutively try to find words that lie between the two words from the previous round.

How we built it

The game platform is a Telegram chat bot that is communicating with multiple possible backends delivering and calculating the word embeddings (word2vec, word2gm, fasttext). The whole stack is written in Python.

Challenges we ran into

• Understanding the structure of embeddings
• finetuning hyperparameters
  • weighting between human and bot input
  • latent representation size
  • dictionary size

Accomplishments that we're proud of

• we built a PoC very fast
• it's actually playable

What we learned

• Word embeddings are less intuitive than one thinks
• "Steinen", "Steine", "Stein" are very similar words

What's next for Play with AI

• other fancy word embeddings j

Built With

• machine-learning
• python
• telegram