Neural Pseudo-Relevance Feedback Models for Sparse and Dense Retrieval

Pseudo-relevance feedback mechanisms have long served as an effective technique to improve the retrieval effectiveness in information retrieval. Recently, large pre-trained language models, such as T5 and BERT, have shown a strong capacity to capture the latent traits of texts. Given the success of these models, we seek to study the capacity of these models for query reformulation. In addition, the BERT models have demonstrated further promise for dense retrieval, where the query and documents are encoded into the contextualised embeddings and relevant documents are retrieved by conducting the semantic matching operation. Although the success of pseudo-relevance feedback for sparse retrieval is well documented, effective pseudo-relevance feedback approaches for dense retrieval paradigm are still in their infancy. Thus, we are concerned with excavating the potential of the pseudo-relevance feedback information combined with the large pre-trained models to conduct effective query reformulation operating on both sparse retrieval and dense retrieval.