RC-Data is a dataset generation framework created by Google DeepMind to produce large-scale reading comprehension question-answer pairs from CNN and Daily Mail news articles. The dataset, introduced in the 2015 paper “Teaching Machines to Read and Comprehend” (Hermann et al., NIPS 2015), was among the first large corpora designed to train and evaluate machine reading and comprehension models. The repository provides scripts for downloading archived CNN and Daily Mail articles from the Wayback Machine and automatically generating cloze-style questions where entities in the text are replaced with placeholders. Each data instance consists of a news article (context), a generated question, and its corresponding answer, making it suitable for supervised machine learning setups. The output follows a standardized question-answer format, with entity mappings to help models resolve named references.
Features
- Generates large-scale question-answer datasets from news articles
- Includes data from CNN and Daily Mail corpora via the Wayback Machine
- Produces questions, contexts, and answers in a standardized text format
- Supports entity anonymization through mapping for model training
- Offers a reproducible generation pipeline using Python scripts
- Compatible with machine comprehension and NLP benchmarking tasks
Project Samples
