Reproduce “Reading Wikipedia to Answer Open-Domain Questions” 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

Danqi Chen, Adam Fisch, Jason Weston, Antoine Bordes

Researchers from Stanford University and Facebook AI Research developed DrQA, a system that answers open-domain factoid questions by first retrieving relevant articles from Wikipedia and then identifying precise answer spans within the text. This approach formalizes the task of Machine Reading at Scale, demonstrating how a modular deep learning architecture can effectively leverage a vast unstructured knowledge source.

Key insights

  • A combination of classical, efficient information retrieval (TF-IDF with bigrams) and deep learning for comprehension forms an effective architecture for open-domain QA on vast corpora.
  • Rich input features, specifically exact match signals and an attention-based aligned question embedding, are crucial for achieving high performance in neural machine comprehension models.
  • Distant supervision effectively generates training data for QA datasets lacking explicit paragraph contexts, and multitask learning across these varied datasets improves model generalization.

Summary by alphaXiv.

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