Enhancing Generative Retrieval Through MultiView Identifier Generation and Fusion

Generative retrieval provides a novel method to information retrieval by generating identifiers based on an input query to match with passages. The quality of these identifiers often determines the performance of generative retrieval. However, existing methods typically over rely on a single identifier, which fails to fully capture the multidimensional semantic information of passages, limiting retrieval performance. This paper proposes a novel multiview identifier framework named MultiView Identifier Generation and Fusion (MIGF), which aims to enhance retrieval by generating various types of identifiers to match information from different positions within a passage. MIGF generates three types of identifiers titles, substrings, and pseudo-queries, and merges them into a more comprehensive new identifier to serve as an additional representation. These identifiers complement to each other, enabling queries to retrieve relevant passages from the perspectives of titles, substrings, and deeper semantic meanings of the passage. MIGF is evaluated on two widely used datasets, TriviaQA and MSMARCO. Experimental results show that our MIGF achieves state-of-the-art performance in generative retrieval, demonstrating its robustness and effectiveness.