Pair Distance Distribution: A Model of Semantic Representation

We introduce PDD (Pair Distance Distribution), a novel corpus-based model of semantic representation. Most corpus-based models are VSMs (Vector Space Models), which while being successful, suffer from both practical and theoretical shortcomings. VSM models produce very large, sparse matrices, and dimensionality reduction is usually performed, leading to high computational complexity, and obscuring the meaning of the dimensions. Similarity in VSMs is constrained to be both symmetric and transitive, contrary to evidence from human subject tests. PDD is featurebased, created automatically from corpora without producing large, sparse matrices. The dimensions along which words are compared are meaningful, enabling better understanding of the model and providing an explanation as to how any two words are similar. Similarity is neither symmetric nor transitive. The model achieved accuracy of 97.6% on a published semantic similarity test.