Neural Retrieval with Partially Shared Embedding Spaces

Abstract

One category of neural information retrieval models tries to learn text representation in a common embedding space for both queries and documents. However, a single embedding space is not always sufficient, since queries and documents are different in terms of length, number of topics covered, etc. We argue that queries and documents should be mapped into different but overlapping embedding spaces, which is named Partially Shared Embedding Space (PSES) model in this paper. PSES consists of two embedding spaces respectively for queries and documents, and a shared embedding space capturing common features of two sources. Those three embeddings are learned by jointly obeying three constraints: a feature separation constraint, a pairwise matching constraint, and a reconstruction constraint. Experiments on standard TREC collections indicate that PSES leads to significant better performance of retrieval over traditional IR models and several neural IR models with only one embedding space.