New York Times' Connections: A Case Study on NLP in Word Games
summary
Builds a pipeline that converts the 16-word board into multi-dimensional semantic graphs using WordNet, ConceptNet and embeddings, scores pairwise relations with a relational alignment heatmap, then generates candidate groupings with a graph convolutional neural network and a reinforcement learner. Random guessing gives ~0% win rate and one-in-6,000 if only one category is known; empirical solve rates cited: ~19% for hard puzzles versus ~70% for easy ones. On a small hard-puzzle subset the GNN+RL system raises solvability noticeably but remains a work in progress.
AI-written summary · about
chapters (11)
- 0:00Introduction and background
- 1:52How Connections game works
- 3:59Why Connections challenges AI
- 7:08Mathematics of random guessing
- 8:12Graph coloring problem explained
- 9:37Semantic similarity limitations
- 11:24Relational alignment in puzzles
- 13:11Semantic distribution evaluation
- 14:54Building semantic graphs
- 15:59Graph neural networks approach
- 17:46Why this research matters
related talks
24mCracking Wordle: Machine Learning based Strategies
34mLearning about AI/ML for Text, with Wordle!
15mAgentic Reasonings with Reinforcement Learning
26mDeep Learning Architectures for Semantic Relation Detection Tasks
56mFrom Go to the GPU, Integrating with CUDA
23mGNN-RAG: Bridging Graph Reasoning and Language Understanding for Knowledge Retrieval
06mCloud-Native GraphRAG: AI’s Logical Edge
11mMutating text toward better nonsense?!