Can prediction-based distributional semantic models predict typicality?

Area of Science:

• Computational Linguistics
• Psycholinguistics
• Cognitive Science

Background:

• Recent computational linguistics advances yield prediction-based semantic models, generating word embeddings (continuous vector representations).
• Psycholinguistic studies explore the psychological plausibility of word embeddings, linking them to various cognitive phenomena like semantic priming and word recognition.

Purpose of the Study:

• To investigate the ability of word embeddings to predict human typicality judgments in category structures.
• To compare the performance of different semantic spaces derived from predictive models against a count-based model.

Main Methods:

• Conducted seven experiments in Dutch and English to predict human typicality judgments for common categories (e.g., birds, fruit, vehicles).
• Extracted predictor variables from various semantic spaces derived from predictive word embedding models.
• Evaluated the predictive power of these variables against established category typicality gradients.

Main Results:

• The performance of predictive word embedding models in capturing human typicality judgments was modest.
• Predictive models did not favorably compare to an older, count-based semantic model.
• Results indicate current predictive models may not fully capture nuanced aspects of human semantic categorization.

Conclusions:

• Despite enthusiasm for predictive models in computational linguistics, their current application in psycholinguistics for category structure prediction shows limitations.
• Further research is needed to explore alternative computational approaches or refine existing models to better align with human semantic cognition.