Reproduce “GLEN: Generative Retrieval via Lexical Index Learning” 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
Sunkyung Lee, Minjin Choi, Jongwuk Lee
Researchers at Sungkyunkwan University developed GLEN, a generative retrieval method that dynamically learns keyword-based identifiers from document content, addressing semantic discrepancy and training-inference gaps in prior approaches. GLEN achieves competitive performance against existing generative retrieval methods and classic baselines by employing a two-phase lexical index learning process and a collision-free inference mechanism.
Key insights
- Learning dynamic lexical identifiers from document content establishes a semantically meaningful connection for generative retrieval.
- A two-phase learning strategy effectively integrates keyword extraction with query-document relevance signals.
- Prefix-aware dynamic negative sampling during training significantly boosts retrieval performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.