This dissertation looks at whether a simple, interpretable scoring function over pre-trained word embeddings can perform well at a cooperative language game, Codenames, and how much the choice of embedding matters. Three research questions are addressed: does risk-aware scoring beat a naive similarity-only baseline, does the choice of embedding affect performance, and does a scoring function calibrated for one embedding still work for others.

The project builds a spymaster agent that works with each embedding behind a shared interface. The scoring function has a reward for hitting target words, and penalties for enemy, neutral, and assassin words together with a threshold that rejects any clue too close to the assassin. The agent is tested through 12,000 simulated games, covering twelve setups across three spymaster strategies (random, naive, and risk-aware) and four embeddings (Word2Vec, GloVe, FastText, and Sentence-BERT). The same set of board seeds is used for every setup to allow for a fair comparison.

The results show that risk-aware scoring outperforms naive scoring on three of the four embeddings, with an increase in win-rate by 18.8 to 29.2 percentage points and decrease in assassin selection rates by 96 to 99 percent. Word2Vec comes out on top, followed by Sentence-BERT and GloVe, and then FastText. The FastText failure traces back to how it builds word vectors from character n-grams, which makes unrelated words look similar to each other, breaking the threshold the scoring function relies on.

These findings suggest that risk-aware scoring functions of this kind do not transfer automatically across embedding families and need to be recalibrated when the underlying similarity geometry changes. The project also shows that simple, transparent designs can produce competitive performances on cooperative language tasks while remaining easy enough to reason with.