In defense of spatial models of semantic representation

Recent semantic space models learn vector representations for word meanings by observing statistical redundancies across a text corpus. A word's meaning is represented as a point in a high-dimensional semantic space, and semantic similarity between words is quantified by a function of their spatial proximity (typically the cosine of the angle between their corresponding vector representations). Recently, Griffiths, Steyvers, and Tenenbaum (2007) demonstrated that spatial models are unable to simulate human free association data due to the constraints placed upon them by metric axioms which appear to be violated in association norms. However, it is important to note that free association data is the product of a retrieval process operating on a semantic representation, and the failures of spatial models are likely be due to mistaking the similarity metric (cosine) for an appropriate process model of the association task-cosine is not what people do with a memory representation. Here, we test the ability of spatial semantic models to simulate association data when they are fused with a simple Luce choice rule to simulate the process of selecting a response in free association. The results provide an existence proof that spatial models can produce the patterns of data in free association previously thought to be problematic for this class of models.