In a paper published today on Arxiv, researchers propose a method that generates drafts of scientific ideas and acts as an automatic research assistant.
PaperRobot, as it was named, is capable of collecting and digesting enormous amounts of human-written papers in some target domain, creating new research ideas and writing some key draft elements of a new research paper.
In their paper, researchers mention that the idea behind PaperRobot was to develop a method that can speed up scientific discovery and production. They identified some key challenges and aspects in the process of scientific investigations and tried to design a method that can overcome those challenges. Some of these challenges include dealing with overwhelming amounts of (potentially related) research papers, generating new ideas by combining different ideas and edge cases, etc.
The proposed framework of PaperRobot consists of few larger modules: background knowledge extraction, link prediction, and new paper writing. They correspond to the three phases of reading and digesting old papers, combining ideas and incrementally improving methods and paper writing.
Researchers leveraged the recent advances in Natural Language Processing (NLP) and Natural Language Understanding (NLU) to design and develop the PaperRobot framework. They used Turing tests to show that the generated paper sections (such as abstracts, conclusion and future work) by PaperRobot were preferred choice over human-written ones. The evaluation showed that human readers prefer PaperRobot generated abstracts more than human-written ones by over 30%, conclusions over 24% and future work sections over 12%.
More details about PaperRobot – the automatic research assistant can be read in the official pre-print paper published on arxiv.