A Semantic Distance based Architecture for a Guesser Agent in ESSENCE's Location Taboo Challenge

Taboo is a word-guessing game in which one player has to describe a target term to another player by giving hints that are neither the target term nor other terms specified in a predetermined list of taboo words. The Location Taboo (LT) Challenge, which has been proposed by the ESSENCE Marie Curie Initial Training Network, is a version of Taboo that only contains cities as target terms and is intended to be played by artificial guesser agents. The hints are extracted from games played by many different human players, whose associations of cities with specific terms are often based on past experiences and therefore very diverse. Modeling this diversity in word associations is one of the main difficulties in solving the LT Challenge. In this paper, we propose a semantic distance based architecture for a guesser agent for the LT Challenge. The proposed architecture employs a two-step approach that narrows down the geographical area of the guess first to the country and then to the city. For ranking countries and cities, different distance metrics are used. As these techniques can be used on web documents crafted by many different individuals, they are well suited to model the diversity in word associations. The results of our evaluation on the LTC test set show that the proposed guesser agent can guess the target city with up to 23.17% accuracy. For 68% of the correct guesses, the proposed agent guesses the target city faster than its human counterpart.