A Siamese BERT based network that plays Codenames
Codenames - or at least this simplified version of it - is a board game in which there are a set of words on the table, some of which correspond to winning words, and other correspond to losing words. Only the hint giver is aware of which words are winning and losing, and they need to give a one word hint to the guesser to figure out which words will win them the game.
In this implementation, a human is the hint giver, and the network is the guesser. In other words, the human player will look at the board and give the neural network a one word hint, which will then give it's best guess of which words you were trying to get it to guess.
For example, if the board looked like this:
Where the green words are winning words, the human player may want to provide a hint such as "butterfly" to ensure the guesser can find the winning words.
A siamese neural network passes two inputs to the same network, and then learns to classify the difference between those two outputs. It is a way to determine whether two words are similar or not, even if our classifier hasn't yet been trained on either of them. This technique is used in a number of real world applications, and is especially prominent in areas like facial recognition.
So in order to play Codenames, our network can run the input word alongside each board word through the siamese network and obtain a similarity score for each pair. From there, it can just pick the words with higher scores, and reject the ones with lower scores.