Innovative Dual-View Training for Enhanced Instruction-Following Retrieval Systems

The field of instruction-following information retrieval (IF-IR) focuses on systems that not only retrieve documents based on relevance but also adhere to specific user instructions, including required features and exclusions. Traditional retrieval models often prioritize semantic relevance, which can lead to overlooking documents that fulfill user instructions. To address this issue, we introduce a dual-view data synthesis method that employs polarity reversal. This technique involves presenting a query alongside a relevant document and a challenging negative example that aligns with the query but does not meet the instruction. By utilizing a large language model (LLM) to generate an alternative instruction that alters the relevance of these documents, we create a training scenario where the same document pair is evaluated under contrasting instructions. This approach compels the retriever to analyze the candidate documents through the lens of the instruction rather than solely relying on topical relevance. Our findings demonstrate a significant 45% performance improvement on the FollowIR benchmark using a 305M-parameter encoder, outperforming other general-purpose models. Additionally, we reveal that data diversity and instruction-based supervision are crucial for enhancing retrieval quality and sensitivity to user instructions.