PFI-TT: Thinking and Talking Books

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

The broader impact of this Partnerships for Innovation - Technology Translation (PFI-TT) project includes the development of generative artificial intelligence (AI)-powered products that unlock the information inside the written documents and books and make them available to any human being. More specifically, the project develops products that allow humans to interact and talk with books and other documents by asking questions in their natural language. The products will collect and organize these human interactions using sophisticated knowledge representation techniques, and use them, along with the informational contents of the books and documents, to power cutting edge generative AI models. Given that some of the most complex tasks and societal challenges, such as education, workforce development, learning, teaching, manufacturing, and research, rely on information, this project could be transformative. The technology could, for example, (1) reduce the cost of education, while also increasing its accessibility, (2) improve the productivity of workers in many industries by allowing them obtain answers to their questions with the click of a button, and (3) improve the speed of scientific progress by helping scientists to remain current with regard to scientific advances. This project combines techniques from several areas of computer science, including formal methods, algorithms, and artificial intelligence, to create the technical foundation for grounding Large Language Models (LLMs) in human knowledge. The project solves an important problem that limits the potential of LLMs - their tendency to "hallucinate" confidently - which make them difficult to use, due to the misleading and potentially harmful information they generate. To this end, the team will develop the technology to represent the informational content of a book (or more generally any document) and discussions pertaining to the book as a knowledge graph. Given the knowledge graph, the team will develop generative AI techniques that will mine the graph to generate the best response to a user query. The techniques work with off-the-shelf LLMs and deliver effective responses by grounding LLMs in the knowledge of human experts embedded in books or human discussions. The project will also develop a web-based platform allowing creation and sharing of knowledge, and engaging in discussions where human beings and generative AI technologies will work together for improved efficiency and scale. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.