Reproduce “Out-of-Domain Semantics to the Rescue! Zero-Shot Hybrid Retrieval Models” 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

Tao Chen, Mingyang Zhang, Jing Lu, Michael Bendersky, Marc Najork

Researchers from Google Research developed zero-shot hybrid retrieval models that combine lexical and deep semantic approaches using Reciprocal Rank Fusion, consistently achieving superior retrieval performance across diverse domains compared to individual models, particularly in out-of-domain settings without any fine-tuning.

Key insights

  • Deep retrieval models, while powerful in-domain, demonstrate significant performance degradation when applied zero-shot to substantially different domains, often yielding poorer results than simpler lexical models.
  • Lexical and deep retrieval models are inherently complementary, as they retrieve distinct sets of relevant documents even when deep models perform weakly out-of-domain.
  • Non-parametric fusion methods like Reciprocal Rank Fusion (RRF) are highly effective for combining heterogeneous retrieval signals in a zero-shot cross-domain context, providing robustness without the need for domain-specific score normalization or weight tuning.

Summary by alphaXiv.

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