The mission of the Allen Institute for Artificial Intelligence (AI2) is to “contribute to humanity through high-impact AI research and engineering.” In support of that mission, the nonprofit — started by Microsoft co-founder Paul Allen — often collaborates with other institutions, and one of its latest partnerships is with the Donald Bren School of Information and Computer Sciences (ICS).
“Researchers from a diversity of institutions bring different perspectives, motivations and previous experiences to a cross-institution collaboration,” says Matt Gardner, a senior research scientist at AI2. “This is a huge help in both identifying and executing on the most promising new research directions.” As a member of the AllenNLP team, Gardner’s research focuses on deep learning for natural language processing, and he’s now furthering that work as a visiting scholar at UCI.
“UCI specifically has a strong history of excellence in machine learning and employs several world-class scientists in the field, making this a really beneficial partnership for us at AI2.” In particular, Gardner is working closely with Assistant Professor of Computer Science Sameer Singh and his NLP group.
“Our high-level research goal is getting machines to read,” says Singh, whose group will host additional AI2 employees and interns throughout the year. In return, AI2 is providing funding for graduate student support. The team is currently working to build AI algorithms that can read and understand paragraph-size text.
“One of Paul Allen’s original goals in starting AI2 was to have a computer that could read a college biology textbook and talk to you about what is in it,” explains Gardner. “Imagine a system that could read all of PubMed and answer doctors’ questions about the latest research related to a particularly challenging case they see in a patient, or a system that could read all of PACER and tell lawyers what precedents are applicable to their current litigation.” The AI2/ICS team is starting with paragraphs but will slowly build toward larger and more complex pieces of text.
In fact, the collaboration has already produced a new research dataset and a related paper, “DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs.” The paper introduces a new English reading comprehension benchmark and a model that combines reading comprehension methods with simple numerical reasoning. Gardner and Singh, along with Singh’s Ph.D. student Dheeru Dua, will present the work in June at the 2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).
“When I moved to Irvine,” notes Gardner, “Sameer and I were thinking of projects to start collaborating on… and this seemed like a natural fit — something I had been wanting to do and that Sameer and Dheeru were excited about working on.”
Singh and his team are indeed excited about the work. “AI2 is a top-tier research institute,” says Singh, pleased with the partnership and looking forward to continued work with AI2 researchers. “We are just now starting to gain momentum.”
Gardner agrees. “We’ve done some great work already, and we have more in progress,” he says. “AI2 is growing in Irvine, and we’re looking forward to a long and fruitful collaboration with UCI.”
— Shani Murray