Improving sparse word similarity models with asymmetric measures by Jean Mark Gawron.
Abstract:
We show that asymmetric models based on Tversky (1977) improve correlations with human similarity judgments and nearest neighbor discovery for both frequent and middle-rank words. In accord with Tversky’s discovery that asymmetric similarity judgments arise when comparing sparse and rich representations, improvement on our two tasks can be traced to heavily weighting the feature bias toward the rarer word when comparing high- and mid- frequency words.
From the introduction:
A key assumption of most models of similarity is that a similarity relation is symmetric. This assumption is foundational for some conceptions, such as the idea of a similarity space, in which similarity is the inverse of distance; and it is deeply embedded into many of the algorithms that build on a similarity relation among objects, such as clustering algorithms. The symmetry assumption is not, however, universal, and it is not essential to all applications of similarity, especially when it comes to modeling human similarity judgments.
What assumptions underlie your “similarity” measures?
Not that we can get away from “assumptions” but are your assumptions based on evidence or are they unexamined assumptions?
Do you know of any general techniques for discovering assumptions in algorithms?