Inspiration The inspiration for Connections AI came from the popular word-categorization puzzle game, similar to The New York Times Connections, where players group words based on thematic similarities. We wanted to bring this experience to an AI model, challenging it to recognize subtle relationships between words, such as synonyms, categories, or concepts. The project merges linguistic insights with AI, creating a fun, interactive way for users to explore how AI understands word relationships.
What it does Connections AI allows users to input a list of words and groups them into sets of four based on semantic similarity. Using the WordNet database and NLP techniques, the AI identifies words that share meanings, themes, or categories, simulating a human-like understanding of language. The model can be tested with various inputs and is designed to handle word categorization tasks with minimal strikes, guiding the user to find correct groups through trial and error.
How we built it We built Connections AI using Python, Flask for the web interface, and NLTK’s WordNet library for word similarity measurements. We used WordNet’s path_similarity metric to assess semantic closeness, grouping words that shared higher similarity scores. Flask manages the backend, handling API requests for user input and returning the AI’s categorized groups, while our custom logic in starter_code.py manages grouping and scoring.
Challenges we ran into One of the biggest challenges was determining the best approach for grouping words by similarity. Since language is nuanced, the model occasionally struggled with words that had multiple meanings or ambiguous relationships. Setting up the Flask API and managing error handling and input/output formats for smooth interaction was another hurdle, as well as ensuring compatibility with different test environments.
Accomplishments that we're proud of We’re proud of implementing a functional model that can group words based on nuanced semantic relationships. Despite the challenges of working with language and ambiguity, we achieved a solution that responds to user inputs, provides meaningful groupings, and evaluates similarity in a way that feels intuitive. Additionally, the seamless API and interactive testing setup represent significant accomplishments in project execution and design.
What we learned Working on Connections AI taught us a great deal about using NLP tools to work with semantic similarity. We learned how to use WordNet in practical applications and gained experience in designing an API that handles multiple input types, manages errors gracefully, and integrates smoothly with a scoring logic. The project highlighted the complexity of language and the need for creative solutions to handle ambiguous or polysemous words.
What's next for Connections AI Next, we’d like to explore using machine learning models trained on language embeddings, like Word2Vec or BERT, to enhance the AI’s ability to recognize more complex relationships between words. Additionally, improving user feedback and scoring metrics to make the game more interactive and adaptable would be a valuable enhancement. We also aim to refine the model to handle larger word sets and more abstract word relationships for a richer and more challenging gameplay experience.
Log in or sign up for Devpost to join the conversation.