Reproduce “SLIDE: Sliding Localized Information for Document Extraction” 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
Divyansh Singh, Manuel Nunez Martinez, Bonnie J. Dorr, Sonja Schmer Galunder
SLIDE (Sliding Localized Information for Document Extraction), a method developed at the University of Florida, addresses LLM limitations in processing long documents and low-resource languages for knowledge graph construction and Retrieval-Augmented Generation (RAG). By intelligently generating and appending localized English context to document chunks, it increased entity extraction by 24% and relationship extraction by 39% for English, and more dramatically for Afrikaans, leading to superior question-answering performance.
Key insights
- Localizing context via a sliding window method allows LLMs to process long documents more effectively, mitigating the 'lost in the middle' phenomenon.
- Generating English context for non-English chunks leverages the LLM's stronger English capabilities, significantly improving knowledge extraction in low-resource languages.
- Enriching document chunks with localized contextual information leads to more comprehensive knowledge graphs, which in turn enhances the performance of downstream Retrieval-Augmented Generation (RAG) systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.