Playing Codenames with Language Graphs and Word Embeddings

Although board games and video have been studied for decades in artificial intelligence research, challenging word remain relatively unexplored. Word are not as constrained like chess or poker. Instead, game strategy is defined by the players’ understanding of way words relate to each other. The Codenames provides a unique opportunity investigate common sense relationships between words, an important open challenge. We propose algorithm that can generate clues from language graph BabelNet any several embedding methods – word2vec, GloVe, fastText BERT. introduce new scoring function measures quality clues, we weighting term called DETECT incorporates dictionary-based representations document frequency improve clue selection. develop BabelNet-Word Selection Framework (BabelNetWSF) overcome computational barriers previously prevented leveraging graphs Codenames. Extensive experiments with human evaluators demonstrate our proposed innovations yield state-of-the-art performance, up 102.8% improvement precision@2 some cases. Overall, this work advances formal study approaches understanding.