Degrees of Separation in Semantic and Syntactic Relationships
Matthew A. Kelly, David Reitter, Robert L West · 2025
Computational models of distributional semantics can ana-lyze a corpus to derive representations of word meanings interms of each word’s relationship to all other words in the cor-pus. While these models are sensitive to topic (e.g., tiger andstripes) and synonymy (e.g., soar and fly), the models havelimited sensitivity to part of speech (e.g., book and shirt areboth nouns). By augmenting a holographic model of semanticmemory with additional levels of representations, we presentevidence that sensitivity to syntax is supported by exploitingassociations between words at varying degrees of separation.We find that sensitivity to associations at three degrees of sep-aration reinforces the relationships between words that sharepart-of-speech and improves the ability of the model to con-struct grammatical sentences. Our model provides evidencethat semantics and syntax exist on a continuum and emergefrom a unitary cognitive system.