The research team under the Department of Computer Science and Engineering in Korea University (KU) once again won the BioASQ competition, an international tournament for artificial intelligence (AI) systems that answer biomedical questions. The team consists of graduate students Jung Min Byeol, Sung Moo Jin, Kim Kang Woo, Yoo Jae Hyo, alumnus Dr. Kim Dong Hyun, and Professor Kang Jae Woo (Department of Computer Science and Engineering).

The BioASQ 8b Phase Challenge presents questions that require answers found in a given paper. After the AI models designed by each competing team take five tests within three months, the results are evaluated by doctors and biologists and announced upon additional review.

The KU team utilized BioBERT, an internationally renowned deep learning language model developed by Professor Kang’s research team that specializes in medical life literature, along with Facebook AI BART. To create more natural sentences for short answer questions, they used BioBERT-NLI (Natural Language Inference), which writes completely new sentences itself. By doing so, the enhanced model replaced the traditional method of selecting sentences from a given document and scored almost perfectly on Readability, an evaluation of how natural a sentence is.

According to the research team, this competition is meaningful in that “the current biomedical question-and-answer system has been improved significantly with the learning of texts within the biomedical domain.” They also added, “Given that an artificial intelligence system can answer questions with sentences natural enough for humans to understand, clinically significant tools that help decision-making are expected to be developed.” Overall, the competition opened up more possibilities for AI innovation in the biomedical field.