-
-
Graph for the word 'a'
-
Graph for the word 'act'
-
Graph for the word 'sing'
Inspiration
Our inspiration came from the New York Times word games and the Elevate App - particularly the crossword games. We wanted to expand on their collection of mentally stimulating mini-games and inspire others to tune their cognitive skills!
What it does
This phone application starts with a letter that is part of a long chain of words. These words are generated from a set of English words and have one letter difference between them. The player needs to deduct the next one-letter-added word with the help of a GPT generated crossword clue to get the next clue. The player wins if they are able to guess all the words in the given chain.
How we built it
Our game couldn't have been made without these three main parts: the graph; the crossword clue generator; the Unity interface.
The graph: The game's main driving force is our one-letter-difference graph. Each node is an English word and has an edge with another node if those words differ by one letter. A subset of the NLTK corpus dataset was cleaned to remove any words that could be unfair for the player to guess - mostly very abstract one or two letter words (e.g, like "tt"). The Python file is able to generate a graph, starting from any word, which shows all possible chains from it. The graph is visualised with networkx and stored as a .graphml file.
The crossword clues: All clues to get from one word to another were generated by GPT-4o-mini through the OpenAI API. This instance of the model was prompt engineered with zero-shot learning so as to not skew the model's responses into limiting its context through giving it examples. The model generates fun, crossword type clues for each word that is used in a chain.
The interface: The frontend and the game logic backend of the game was made in Unity in the aspect ratio of a smartphone. A word chain is randomly selected upon starting the game and its corresponding clues are displayed on the screen. The player can drag and insert letters from keys from a keyboard and create a new word. If the new word is the word that corresponds with the clue, the next clue displays for this word.
Challenges we ran into
Creating the graph was a challenge as we had to deal with 900 million possible edges to begin with. Working out how to narrow this down to a computable size and traverse the graph were challenges we were able to overcome!
Accomplishments that we're proud of
Creating the word graph and using OpenAI API for the crossword generation! To be able to achieve their levels of functionality and success is a huge achievement for us. Forming the algorithm to discover the longest chain in our graph was another proud moment.
What we learned
Fully understanding eachother's collaboration and working style in such a pressured environment. We now also know that we are both able to work into the night with determination to finish something!
What's next for skrobl
New York Times will certainly buy us (!). It was an idea that we have wanted to develop for a long time and with this incredible start we can continue to develop the application. From brain storming ideas and discussions with others, other game modes are on the cards. Modes such as removing a letter each round and spiral crossword are next!
Log in or sign up for Devpost to join the conversation.