Learning to Re-rank for Interactive Problem Resolution and Query Refinement

We study the design of an information retrieval (IR) system that assists customer service agents while they interact with end-users. The type of IR needed is difficult because of the large lexical gap between problems as described by customers, and solutions. We describe an approach that bridges this lexical gap by learning semantic relatedness using tensor representations. Queries that are short and vague, which are common in practice, result in a large number of documents being retrieved, and a high cognitive load for customer service agents. We show how to reduce this burden by providing suggestions that are selected based on the learned measures of semantic relatedness. Experiments show that the approach offers substantial benefit compared to the use of standard lexical similarity.