Modelling Word Associations with Word Embeddings for a Guesser Agent in the Taboo City Challenge Competition

In the Taboo City Challenge, artificial agents should guess the names of cities from simple textual hints and are evaluated with games played by humans. Thus, playing the games successfully requires mimicking associations that humans have with geographical locations. In this paper, an architecture is proposed that calculates the associative similarity between a city and a hint from a semantic vector space. The semantic vector space is created using the Skip-gram hierarchical softmax model, from a tailored corpus about travel destinations. We investigate the effect of varying training parameters and introduce a targeted corpus annotation method that significantly improves performance. The results on a dataset of 149 games indicate that the proposed architecture can guess the target city with up to 22.45% accuracy — a substantial improvement over the 4.11% accuracy achieved by the baseline architecture.