Reproduce “Beyond Matching: Category-Guided Latent Intent Reasoning for Generative Retrieval in E-Commerce” 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

Fuwei Zhang, Xiaoyu Liu, Jiajie Jin, Jiale Mao, Wei Chen, Dongbo Xi, Yifan Yang, Peng Yan and 3 more

CaLIR, a generative retrieval framework from Beihang University and Meituan, enhances e-commerce search by employing category-guided latent intent reasoning to align natural language queries with abstract product Semantic Identifiers (SIDs). This method improved Recall@100 by over 10% on multilingual ESCI datasets compared to leading baselines, achieving semantic grounding without significant latency increases.

Key insights

  • Internalizing reasoning within continuous latent states, guided by hierarchical category structures, can bridge the "representation gap" more efficiently than explicit textual Chain-of-Thought for online generative retrieval.
  • Explicitly supervising latent states with product category hierarchies and regularizing them to capture multiple positive intents enhances the model's understanding of complex e-commerce queries.
  • Leveraging inferred category information to dynamically constrain the SID decoding search space significantly improves retrieval precision without introducing prohibitive latency for online systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.