Complex information retrieval tasks often require selecting not just the most relevant items, but a set of objects whose internal structure and mutual relationships align with a natural language query. Standard dense and sparse retrievers often focus on surface-level similarity and overlook relationships among individual objects, leading to suboptimal ranking. Recent work on the Alignment-Oriented Retrieval Method (ARM) demonstrates that structured alignment and joint selection can improve open-domain question answering, but its applicability to entity-centric search has not been explored. In this study, we introduce S-ARM (Structured Alignment-Oriented Retrieval Model), an adaptation of ARM designed for structured object retrieval. S-ARM extends alignment principles beyond text passages to handle LinkedIn-style professional profiles composed of heterogeneous fields such as name, headline, education, and experience. We evaluate S-ARM on a corpus of 10,000 structured profiles using a benchmark of 492 human-verified natural language queries generated via an LLM-assisted synthesis process and multi-annotator validation. Our findings show that S-ARM improves object-level retrieval quality across both dense (embedding-based) and sparse (lexical) backbones, achieving a consistent trade-off between effectiveness and efficiency.
Owais Talaat Waheed (Thu,) studied this question.