AI2 (Allen Institute for Artificial Intelligence) has recently released the world’s largest open dataset for training language models, in an effort to make natural language processing (NLP) models more efficient and accurate. The dataset, which includes more than 9 million webpages and 700 million words, is three times bigger than the previous largest open dataset.
The new dataset is called the AI2 Reasoning Challenge (ARC). It consists of more than 8,000 common core-aligned 8th grade educational questions, all of which are factoid questions such as “What is the capital of California?” The dataset has both multiple-choice and fill-in-the-blank formatted questions with potential answers. This allows scientists to use the AI2 dataset to better develop new learning systems and improve existing models on NLP tasks.
The AI2 dataset was specifically developed for competition use. The first competition to use the ARC dataset was the AI2 Reasoning Challenge (ARC) held at the NeurIPS conference in 2018. At the competition, AI2 challenged attendees to build a system that could correctly answer all the questions in the dataset. Researchers came from multiple organizations and universities and used various deep learning techniques to compete. The competition spurred the development of multiple systems and approaches towards tackling the challenge.
AI2 hopes that the ARC dataset will be widely used in the research community and help with the development of better, more efficient, and more accurate NLP models. AI2 also plans to use the dataset to develop open-source libraries that can be used by organizations to build their own NLP systems. This large open dataset can be used to train systems to analyze and comprehend text more effectively and accurately, which has the potential to revolutionize the field of natural language processing and lead to more accurate machine-readers.
