Free Associations as Steady States in Dynamic Spaces

The free association task provides a glimpse into the organizational structure of concepts in memory, and has been used by theorists as a benchmark for computational models of semantic processing. The most successful match to free association data comes from a high-level probabilistic approach called Topics modeling (Griffiths, Steyvers, & Tennenbaum, 2007). Topics models treat concepts as probability distributions over a set of discrete themes, or topics. To date, there has been no successful process-model of the free-association data-set. One path forward is to use high-dimensional vector representations as items in a broader control architecture, however, such approach decouples the encoding process from retrieval. We offer an alternative approach using a Dynamic Eigen Net (DEN), an associative net with dynamic eigenvectors, as a way to simultaneously model encoding and retrieval. We start with a replication of the Topics model of free-association, ensuring our corpus is matched to the corpus they used in their simulations. We then train a DEN, along with three other popular semantic representation algorithms. Our results show that a DEN provides a similar match to data as to the Topics model, while providing a better match than the three other algorithms. We suggest DENs as possible process-model compliments to the probabilistic account.