Reproduce “AutoMIR: Effective Zero-Shot Medical Information Retrieval without Relevance Labels” 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

Lei Li, Xiangxu Zhang, Xiao Zhou, Zheng Liu

Researchers at Renmin University of China developed SL-HyDE, a self-learning framework for zero-shot medical information retrieval that operates effectively without relevance labels. This method, which mutually enhances a generator and a retriever, achieved an average NDCG@10 improvement of 4.87% over baseline HyDE on the newly introduced Chinese Medical Information Retrieval Benchmark (CMIRB).

Key insights

  • A self-learning framework can effectively adapt both large language models (LLMs) and dense retrievers to specialized domains like medicine by iteratively generating optimal hypothetical documents and training on internal supervisory signals from unlabeled corpora.
  • Fine-tuning both the generator and the retriever components is essential for maximizing performance gains in zero-shot medical information retrieval, with the retriever's fine-tuning contributing significantly.
  • The fusion strategy of query and hypothetical document embeddings is critical, with mean pooling demonstrating superior performance compared to relying solely on hypothetical documents or string concatenation.

Summary by alphaXiv.

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