Robust Word Similarity Estimation Using Perturbation Kernels

We introduce perturbation kernels, a new class of similarity measure for information retrieval that casts word similarity in terms of multi-task learning. Perturbation kernels model uncertainty in the user’s query by choosing a small number of variations in the relative weights of the query terms to build a more complete picture of the query context, which is then used to compute a form of expected distance between words. Our approach has a principled mathematical foundation, a simple analytical form, and makes few assumptions about the underlying retrieval model, making it easy to apply in a broad family of existing query expansion and model estimation algorithms.