A Formal Product Search Model with Ensembled Proximity

In this paper we study the problem of product search, where products are retrieved and ranked based on how their reviews match the query. Current product search systems suffer from the incapability to measure correspondence between a product feature and its desired property. A proximity language model is presented to embed textual adjacency in the frequency based estimation framework. To tailor for product search problem, we explore strategies for distinguishing product feature and desired property, quantifying pair-wise proximity based on conditional probability, and aggregating review opinions at product level. Experiments on a real data set demonstrate good performances of our model.