Inspiration The inspiration for Connections AI came from the desire to recreate the classic word association game, where players group words based on shared meanings or themes. We wanted to build a challenging single-player experience using AI that simulates human intuition, creating an engaging gameplay loop that feels satisfying to solve. By leveraging natural language processing (NLP) techniques, we aimed to capture the essence of associative thinking and push the boundaries of word clustering through machine learning.
What it does Connections AI presents players with a grid of 16 words, challenging them to identify groups of four words that share a common category or theme. Players have four lives and receive feedback when they get three out of four words correct, allowing them to refine their guesses. The goal is to correctly group all words, with the AI adapting its suggestions based on player feedback to improve its categorization accuracy.
How we built it We used the Sentence-BERT model to generate embeddings for the words, capturing their semantic relationships. By applying Agglomerative Clustering, we grouped words into clusters of related meanings, and implemented a refined “One Word Away” strategy to provide context-aware hints. Python and the sentence-transformers library formed the foundation, with additional logic to handle game mechanics and track player feedback.
Challenges we ran into One of our main challenges was achieving high accuracy in clustering. Some words were difficult to categorize due to overlapping meanings, leading to frequent incorrect guesses. Fine-tuning the model and clustering strategy to prevent repeated incorrect guesses required iterative adjustments and experimentation with different embedding models and clustering methods.
Accomplishments that we're proud of We’re proud of creating a dynamic AI that provides an engaging gameplay experience while learning from player interactions. Our refined “One Word Away” mechanism improves the AI's adaptability, and we’ve managed to make meaningful improvements in clustering performance, ensuring the AI suggests plausible groups even with complex word associations.
What we learned Building Connections AI taught us a lot about handling NLP tasks in a game context. We gained hands-on experience in using sentence embeddings for semantic clustering and refined our knowledge of game design mechanics that keep players engaged. We also learned the importance of balancing AI accuracy with randomness to maintain an enjoyable game flow.
What's next for Connections AI Next, we aim to improve Connections AI by incorporating advanced word embedding models and contextual fine-tuning to enhance accuracy. We plan to add new difficulty levels and additional feedback mechanisms to help players identify group themes. Expanding the game’s word database and refining our clustering algorithms will ensure it continues to challenge players with fresh and varied content.
Log in or sign up for Devpost to join the conversation.