We present a recommender system intended to be used by a community of gamers. The system uses free-form text reviews of games written by the members of the community, along with information about the games that a particular user likes, in order to recommend new games that are likely to be of interest to that user. The system uses the frequency of co-occurrence of word pairs that appear in the reviews of a game as features that represent the game. The pairs consist of adjectives and context words; i.e., words that appear close to an adjective in a review. Because of the extremely large number of possible combinations of adjectives and context words, we use information-theoretic co-clustering of the adjective-context word pairs to reduce the dimensionality. Games are represented using the standard information retrieval vector space model, in which vector features are based on the frequency of occurrence of cocluster pairs.We present the results of three experiments with our system. In the first experiment, we use a variety of strategies to relate frequencies of co-cluster pairs to vector features, to see which produces the most accurate recommendations. In the second, we explore the effects of co-cluster dimensionality on the quality of our system’s recommendations. In the third experiment, we compare our approach to a baseline approach using a bag-of-words technique and conclude that our approach produces higher quality recommendations.