Date of Award


Degree Type


Degree Name



Computer Science

First Advisor

Richard Caron


Applied sciences, Psychology, Distributional semantics, Embodied cognition, Knowledge representation, Neural network, Semantics




The symbol interdependency hypothesis (Louwerse, 2007, 2008) posits that word meaning is dependent upon two sources of information: embodied or grounded knowledge, obtained from observation of and interaction with the physical world, and symbolic or co-occurrence information, gleaned from experience with how words are used together in written and spoken language. This theory assumes that embodied properties of objects influence the statistical structure of language to such an extent that the embodied properties become partially encoded within the structure of language. The work presented in this dissertation provides support for the symbol interdependency hypothesis by demonstrating that grounded knowledge (in the form of physical and behavioural properties of living and non-living objects) can be identified by analyzing word usage in a large body of written text. An automated method of creating high-dimensional vector-based semantic representations is presented. Several demonstrations show that the representations capture word meaning in a way that aligns with intuition and are able to reproduce non-intuitive results of experiments from the psycholinguistic literature. A feedforward neural network was trained to produce a list of physical and behavioural properties of an object in response to the object's high-dimensional vector representation. The resulting network was able to identify features of the concepts on which it was trained with near-perfect accuracy and was able to generalize this ability to novel concepts and identify properties of concepts to which it was not previously exposed. These results indicate that there is sufficient information in word usage to identify embodied properties of concepts, a finding that is consistent with the symbol interdependency hypothesis.