Information Retrieval Re-Ranking Method Based on Bidirectional Text Expansion

With the rapid development of large language models (LLMs), remarkable progress has been made in both text matching and text expansion technologies in information retrieval. As two important methods for enhancing text representation, query expansion and document expansion have been widely applied in modern information retrieval systems. Currently, mainstream text expansion methods primarily rely on large language models. However, there are obvious differences in language diversity and style between the text generated by these models and the text created manually. These differences may affect the accuracy of calculating the query-document relevance, ultimately leading to a decline in the performance of the entire information retrieval system. To address this issue, this paper proposed an information retrieval re-ranking method based on bidirectional text expansion (BTE-IRRM). First, zero-shot prompting was used to enable the large language model to generate pseudo-queries for documents and pseudo-documents for queries. Then, the semantic similarity between these pseudo-queries and pseudo-documents was calculated. Finally, the similarity scores of the original query-document and the semantic similarity scores of the pseudo-query-pseudo-document were weighted and fused to obtain the final document ranking result. To validate the effectiveness of the proposed method, experiments were conducted on two public datasets (DL19 and DL20). Experimental results demonstrate that compared with the existing baseline methods, the BTE-IRRM method has achieved significant improvements in multiple evaluation indicators. Therefore, the bidirectional text expansion method proposed in this paper can further enhance the relevance matching between queries and documents, thereby improving the performance of the entire information retrieval system.