Organizing the space and behavior of semantic models.

Abstract

Semantic models play an important role in cognitive science. These models use statistical learning to model word meanings from co-occurrences in text corpora. A wide variety of semantic models have been proposed, and the literature has typically emphasized situations in which one model outperforms another. However, because these models often vary with respect to multiple sub-processes (e.g., their normalization or dimensionality-reduction methods), it can be difficult to delineate which of these processes are responsible for observed performance differences. Furthermore, the fact that any two models may vary along multiple dimensions makes it difficult to understand where these models fall within the space of possible psychological theories. In this paper, we propose a general framework for organizing the space of semantic models. We then illustrate how this framework can be used to understand model comparisons in terms of individual manipulations along sub-processes. Using several artificial datasets we show how both representational structure and dimensionality-reduction influence a model's ability to pick up on different types of word relationships.