Factorized Hypothesis Search for Evidence-to-Taxonomy Retrieval
Linhai Ma, Ethan F. Wei, Xueqing Peng, Yan Wang, Lingfei Qian, Víctor Gutiérrez-Basulto
cs.CL, cs.AI
2026-08-07
FHS argues taxonomy retrieval is bottlenecked by the query, not the index; factorizing evidence into competing hypotheses gives the best non-oracle Recall@1 on finance tagging and clinical coding.
Mapping an observed input to an entry in a large concept inventory is everywhere: finance has the 17,388-concept US-GAAP taxonomy, clinical coding has the 71,344-code ICD-10-CM. The standard retrieve-and-rerank approach works when the input directly states the target concept. In practice the input is often indirect evidence: a table cell whose meaning depends on its row, column, datatype, and context, or a clinical mention whose sense depends on the symptom, specimen, and surrounding course. The paper calls this mismatch the retrieval readiness gap.
An oracle probe quantifies the gap sharply. Querying with each gold concept's own label plus definition hits the top ten for 100% of facts, but raw evidence retrieval places the gold in the top 200 for 75.2% of facts and in the top ten for only 24.0%. The conclusion is blunt: the index is not the bottleneck, the query is. The target is usually reachable but ranked deep. The natural fix of rewriting the context into a single query (HyDE-style) assumes the semantics are certain and only need lexical polish, whereas here the semantics themselves are ambiguous: one fact often has several plausible readings, and committing too early steers retrieval toward the wrong sub-hierarchy.
FHS maintains multiple partial interpretations over named semantic dimensions, called factorized hypotheses. Each assigns values on six dimensions (Family, Role, Event, Qualifier, Scope, Temporal), leaving a dimension blank when the evidence does not support it rather than guessing. Blanks are free: the renderer skips them and scoring excludes them, while a wrong guessed value enters the query and drags retrieval toward an unsupported region.
Within one generation call, each hypothesis produces two queries: a label-form query built programmatically by concatenating resolved dimension values under the same tokenization as the index labels, and a definition-form query written as natural language by the LLM. Rankings across all hypotheses are fused with reciprocal rank fusion into a candidate pool. A candidate-level verifier then takes a structurally diverse window per hypothesis and asks the LLM for support, no-support, or abstain on each dimension for each candidate, combining support with the retrieval score to reorder the head. The factorized representation doubles as the shared schema for both query generation and candidate verification, which is what separates it from unstructured free-text sampling.
Several choices matter. The generator is sampled unmodified rather than forced to differ, because forcing diversity lowers each hypothesis and reduces coverage. A single parallel round suffices: an FHS-Seq control that adds sequential refinement changes the top-50 pool substantially but does not improve Recall@50 and costs much more. And the definition form matters far more than the label form (removing it drops Recall@1 by 0.034, removing the label form barely moves it).
On financial tagging (2024 US-GAAP), FHS reaches Recall@1 of 0.185, ahead of the strongest baseline (retrieval-feedback refinement at 0.141) by 0.044; MRR rises from 0.238 to 0.257 and final accuracy from 0.234 to 0.255. On CodiEsp clinical coding (ICD-10-CM), Recall@1 goes from the best baseline's 0.201 to 0.264, MRR from 0.298 to 0.352, and accuracy from 0.322 to 0.330. FHS is the strongest head-ranking method in both domains.
Deeper in the ranking the story is nuanced. On financial tagging, retrieval-feedback refinement reaches Recall@50 of 0.660, above FHS's 0.543: iteration enlarges the candidate set but buys neither a better head nor a better final prediction. Ablations confirm this. Replacing the factorized path with a same-size free-text ensemble drops Recall@1 from 0.185 to 0.116 and MRR from 0.257 to 0.194, yet Recall@50 rises to 0.583. Factorization trades for precision, not coverage.
For any task that maps indirect evidence onto a large concept inventory (financial XBRL tagging, clinical ICD coding, schema matching, entity linking), FHS offers a clear diagnostic lens: separate whether the bottleneck is the index or the query. Its answer is to let the model hold several competing interpretations at once and have candidates justify themselves against each, rather than polishing a single query. Code and data are public.
The authors concede that in both domains the semantic dimensions are derived from the existing taxonomy's structure and metadata, so applying FHS to an inventory without such structure requires defining or inducing a schema, with automated induction left to future work, and that all LLM components use Qwen3-32B, so variation across model families and scales is untested. Two more from reading: absolute Recall@1 is still modest (0.185, 0.264) on a genuinely hard task where iterative methods are stronger at deep recall, and the oracle-choice upper bound sits only slightly above FHS (Recall@1 0.222), which both explains why sequential refinement does not help and shows the hypothesis generator still has room to improve.