Bi-Phase Enhanced IVFPQ for Time-Efficient Ad-hoc Retrieval

IVFPQ is a popular index paradigm for time-efficient ad-hoc retrieval. Instead of traversing the entire database for relevant documents, it accelerates the retrieval operation by 1) accessing a fraction of the database guided the activation of latent topics in IVF (inverted file system), and 2) approximating the exact relevance measurement based on PQ (product quantization). However, the conventional IVFPQ is limited in retrieval performance due to the coarse granularity of its latent topics. On the one hand, it may result in severe loss of retrieval quality when visiting a small number of topics; on the other hand, it will lead to a huge retrieval cost when visiting a large number of topics. To mitigate the above problem, we propose a novel framework named Bi-Phase IVFPQ. It jointly uses two types of features: the latent topics and the explicit terms, to build the inverted file system. Both types of features are complementary to each other, which helps to achieve better coverage of the relevant documents. Besides, the documents' memberships to different IVF entries are learned by distilling knowledge from deep semantic models, which substantially improves the index quality and retrieval accuracy. We perform comprehensive empirical studies on popular ad-hoc retrieval benchmarks, whose results verify the effectiveness and efficiency of our proposed framework.