Reproduce “CoQA: A Conversational Question Answering Challenge” 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
Siva Reddy, Danqi Chen, Christopher D. Manning
A large-scale dataset for conversational question answering, CoQA, features 127k question-answer pairs from diverse domains, requiring models to resolve coreferences and generate free-form responses. Initial machine models achieved 65.4% F1, leaving a 23.4 F1 point gap to human performance of 88.8%.
Key insights
- Nearly 70% of questions in CoQA depend on conversation history through explicit or implicit coreference, highlighting the critical role of context in multi-turn QA.
- A significant portion of answers (33.2%) are free-form, including "Yes/No" and paraphrases, demonstrating the need for abstractive capabilities beyond direct text extraction.
- Questions requiring pragmatic reasoning and those without coreference proved more challenging for models, and all models exhibited a notable performance drop (6.6 F1 points) on out-of-domain data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.