Can prediction-based distributional semantic models predict typicality?

Recent advances in the field of computational linguistics have led to the development of various prediction-based models of semantics. These models seek to infer word representations from large text collections by predicting target words from neighbouring words (or vice versa). The resulting representations are vectors in a continuous space, collectively called word embeddings. Although psychological plausibility was not a primary concern for the developers of predictive models, it has been the topic of several recent studies in the field of psycholinguistics. That is, word embeddings have been linked to similarity ratings, word associations, semantic priming, word recognition latencies, and so on. Here, we build on this work by investigating category structure. Throughout seven experiments, we sought to predict human typicality judgements from two languages, Dutch and English, using different semantic spaces. More specifically, we extracted a number of predictor variables, and evaluated how well they could capture the typicality gradient of common categories (e.g., birds, fruit, vehicles, etc.). Overall, the performance of predictive models was rather modest and did not compare favourably with that of an older count-based model. These results are somewhat disappointing given the enthusiasm surrounding predictive models. Possible explanations and future directions are discussed.