■ The AI program developed by LG Sciencepark came in first in the machine reading comprehension evaluation of KorQuAD 1.0, a Korean standard dataset for AI learning
■ The company plans to apply the technology to customer chatbot service
■ It also topped the world's most prestigious AI contest last month
LG has ranked first in the evaluation of artificial intelligence (AI) machine reading comprehension of the Korean language.
The AI program developed by LG Sciencepark came in first with 95.39 points in the machine reading comprehension evaluation of KorQuAD 1.0, a Korean standard dataset for AI learning. This is a higher result than the score (91.2 points) of humans who solved the same reading problem.
Machine reading comprehension is technology in which AI analyzes problems by itself and finds answers optimized for questions by repeatedly learning to "question and answer."
For example, if given the text "Jeju-do is an island located on the South Sea of Korea and is the country’s largest island. It belongs to the administrative district of Jeju Special Self-Governing Province. The area of the island is 1833.2 km², which accounts for 1.83% of the area of South Korea. In 2020, the number of registered residents was about 670,000, becoming the most populous of all islands in South Korea" and the question "How many people do live in Jeju-do?," the AI understands the context and answers "670,000 people."
Reading evaluation is conducted in a way that AI answers thousands of atypical, short-answer questions for about 10 minutes, and scores are given according to the levels of precision and reproduction of the answers.
LG plans to apply this technology to its customer service chatbot in the future. An advanced chatbot can understand the meaning of the question and provide natural answers like a real person, unlike existing chatbots that respond by selecting one among the set answers according to the type of questions.
This is because AI technology can quickly and accurately comprehend a vast amount of long and complex documents such as product manuals and web pages in Korean and find necessary key information in a short time.
LG Sciencepark’s AI Division, which made this accomplishment, also topped the Continual Learning Challenge held by the 2020 Computer Vision and Pattern Recognition (CVPR), the world's most prestigious computer vision society, last month.
Continual learning, a technology that improves accuracy by learning new tasks without forgetting the previously learned tasks, is considered one of the most difficult challenges in the AI field. LG Sciencepark participated as a joint research team with the University of Toronto and won the competition among a total of 79 teams including Amazon, the Chinese Academy of Sciences, and the University of Tokyo.
LG Sciencepark’s AI Division is a dedicated AI organization within the group that sets AI strategies and conducts research on existing technologies.
"We have proven LG's unrivaled AI technology by consecutively winning first place in domestic and international AI evaluations," said Kyung-hoon Bae, head of the AI Division at LG Sciencepark. "In order to secure global competitiveness in the AI field, we will increase investment in R&D, boldly promote open innovation, and actively recruit talent."