A US government research lab has built a tool for finding and summarising knowledge in scientific paper abstracts. It uses the same AI that powers the recently publicly released chatbot ChatGPT.
鈥淭here is a line of sight to the time when we will have research assistants that are AIs that have access to incredible amounts of information,鈥� says .
Dagdelen and his colleagues at Lawrence Berkeley National Laboratory in California started out with the goal of teaching GPT-3 鈥� an AI built by the company OpenAI that can spit out text summaries and answer conversational prompts 鈥� to read the abstracts of materials science research papers, then find relevant information and present it in a useful way.
Advertisement
First, the team showed the AI 100 examples of this task 鈥� for example, taking an abstract about a lithium battery with experimental chemical components and presenting the information in a standard data format that clearly identifies the chemical names, formulas, general descriptions and practical applications.
Having shown the AI what to do, the researchers used it to help generate 500 more examples. The team then corrected any mistakes and fed the examples back into the AI to fine-tune its performance.
The team then tested it on 65 abstracts and found that it could pick out about 75 per cent of the relevant information in any given abstract.
Any researcher could follow the approach of feeding GPT-3 examples of how they would like the AI to retrieve and format scientific knowledge from abstracts. But one limitation is that GPT-3 and other AI services limit the amount of text that can be fed into them, which means researchers cannot easily use GPT-3 to analyse entire research papers, says , part of the team.
AIs that can accurately extract information from text would represent a 鈥減aradigm shift鈥� in helping researchers access scientific knowledge, says at the University of Cambridge in the UK.
GPT-3 is known to produce mistakes and that was also the case here. The best fixes seem to be teaching the AI using corrected examples of its worst mistakes, or else teaching the AI to extract data exactly as it originally appears in the abstracts, says team member .
This strategy for teaching GPT-3 to sift through existing research for useful nuggets of knowledge could also work for other scientific fields beyond just materials science. 鈥淏efore, some academic researcher was going to be combing through hundreds or thousands of papers,鈥� says Dunn. 鈥淣ow we can have AI go through 1000 times or more and really get a comprehensive picture of the data in a particular subdomain of science.鈥�
Reference:
