Reproduce “Learning Multiresolution Relevance for Hierarchical Generative Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weihao Shen, Wei Chen, Fuwei Zhang, Guojun Liu, Qingsong Hua, Wei Lin, Fuzhen Zhuang
RARS aligns relevance supervision with the successive decisions of hierarchical generative retrieval by turning document-level relevance into conditional distributions over branches of semantic identifier trees. On three ESCI product-search locales, it improves Recall@100 and NDCG@10 over matched training baselines, while its auxiliary predictor is not needed at inference.
Key insights
- Relevance can be modeled as a distribution over successive refinements: mass on a parent prefix equals the combined mass of its children, making supervision coherent across depths.
- Observed ambiguity is common near the top of the hierarchy: roughly 70–80% of eligible queries branch at the first level, compared with fewer than 20% at the fourth.
- Incomplete relevance judgments can hide real branching, so the method’s targets reflect available annotations and may not capture the full relevance distribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.