Factorized Hypothesis Search Improves Evidence-to-Taxonomy Retrieval

A recent preprint on arXiv (2608.06614) presents Factorized Hypothesis Search (FHS), a novel approach aimed at bridging the 'retrieval readiness gap' in large-taxonomy retrieval. This gap arises when input data, such as table cells, lacks clear semantic meaning and relies on contextual elements like row, column, and datatype. The authors introduce FHS to preserve various partial interpretations across named semantic dimensions, facilitating structured query rendering, multi-hypothesis retrieval, and dimension-level candidate verification. In assessments involving financial taxonomy tagging and CodiEsp clinical coding tasks, FHS outperforms non-oracle methods in Recall@1, MRR, and overall accuracy. Significantly, substituting the factorized hypothesis path with a free-text ensemble leads to the most considerable decline in performance, highlighting the factorized method's value. This paper marks a notable progress in retrieval systems for intricate, context-sensitive evidence.