Factorized Hypothesis Search improves large-taxonomy retrieval on financial and clinical tasks

TheFinAI · hf · 2026-08-11

TheFinAI proposes Factorized Hypothesis Search (FHS) to address the retrieval readiness gap in large-taxonomy retrieval, where indirect evidence (e.g., table cells) depends on context. FHS maintains multiple partial interpretations over semantic dimensions, enabling structured query rendering, multi-hypothesis retrieval, and dimension-level verification. On financial taxonomy tagging and CodiEsp clinical coding, FHS achieves best Recall@1, MRR, and accuracy among non-oracle methods. Ablations show replacing factorized hypotheses with free-text ensemble causes largest drop in head-ranking performance, while sequential refinement adds no gain over FHS's strong parallel first round.

Original post →

More from Research

Research channel →