Prof. Jaewoo Kang’s team succeeds in AI reading and learning
Higher level of understanding than domain experts in biomedical literature
Results published in JMIR Medical Informatics
▲ Prof. Jaewoo Kang’s team (clockwise from top left: Prof. Jaewoo Kang; Dr. Seongsoon Kim; Yonghwa Choi, integrated MS-PhD student; Donghyeon Park, doctoral student)
A research team led by Professor Jaewoo Kang of the Department of Computer Science and Engineering succeeded in reading comprehension and knowledge acquisition using artificial intelligence.
With the recent increase in text-based information, there has been a surge in machine comprehension research in the field of AI and natural language processing. In machine comprehension, machines are trained to read and understand a given text, and then asked questions to test their understanding.
Existing research has mainly focused on less difficult texts such as newspaper articles and storybooks. The question of whether machines are capable of understanding biomedical papers and other texts requiring expert-level knowledge had not been addressed.
Prof. Jaewoo Kang’s team expanded the existing deep-learning machine comprehension system to the biomedical domain. The model was trained to read tens of thousands of biomedical articles and acquire expert knowledge on its own, and then provide answers to related questions. In addition, it was provided with detailed information on biomedical entities appearing in the datasets.
The AI model was tested by providing abstracts of biomedical articles and asking related questions. The machine comprehension model developed by Professor Jaewoo Kang achieved an accuracy of 92% in cancer-related questions, while human experts (faculty at the University of Colorado Anschutz Medical Campus) scored only 66%. Unlike other AI models that showed similar performance to humans, the proposed model fared much better than humans.
The machine comprehension model outperformed human experts not only in accuracy, but also in speed. The machine comprehension system took less than 0.1 seconds (0.06 seconds) to solve a total of 50 questions, while human experts needed close to two hours (115 minutes).
In the coming age of precision medicine, the ability to rapidly acquire new knowledge and make inferences will be the key to international competitiveness. Machine comprehension systems capable of absorbing newly generated knowledge (more than 3,000 papers published per day in the biomedical field), in addition to the already massive database of knowledge, can serve human experts as decision-making tools.