Kakao Enterprise Achieves No.1 in Korean Machine Reading Comprehension Evaluation with Proprietary AI Model

[Asia Economy Reporter Seungjin Lee] Kakao Enterprise announced on the 23rd that its self-developed natural language processing artificial intelligence (AI) model ranked first in the performance evaluation of the Korean machine reading comprehension dataset ‘KorQuAD 2.0 (The Korean Question Answering Dataset 2.0)’.

The Korean question answering dataset 'KorQuAD' is a machine reading comprehension ability evaluation created by LG CNS in 2018. Since the open-source release of LGCNS’s ‘KorQuAD 1.0’ in December 2018, the current 2.0 version requires technologies for understanding HTML documents containing lists and tables, as well as quickly processing long and extensive documents, thus increasing the difficulty level.

The evaluation is conducted by presenting questions extracted from the dataset, and the AI model finds the correct answers. Domestic companies such as Samsung SDS and Naver Clova participated in the evaluation to test their AI model performance.

Kakao Enterprise’s natural language processing model 'LittleBird-large' recorded an accuracy score of 90.22 points, ranking first. Not only accuracy but also processing speed demonstrated technological prowess by being approximately 1.7 times faster than the previous top model. Additionally, the ‘LittleBird-base’ model, developed together by optimizing model size to increase inference speed and enhance practicality, ranked sixth with an accuracy of 88.57 points, achieving data processing speeds about 3.4 times faster than similar performance models from other companies.

Kakao Enterprise plans to continuously advance these models and apply them to search, chatbots, and the AI contact center ‘Kakao i Connect Center.’ For example, when applied to the ‘Kakao i Connect Center,’ the AI counselor can utilize the knowledge search system based on this machine reading comprehension model to provide more accurate answers to customer inquiries and requests.

In 2019, Kakao’s own language model also ranked first in 'KorQuAD 1.0.' At that time, Kakao’s language model scored 91.85 points, 0.65 points higher than the human sentence discrimination and reading comprehension ability level of 91.20 points, marking the first case where AI scored higher than humans in the evaluation.

Choi Dong-jin, Head of AI Lab & Service at Kakao Enterprise, stated, “Machine reading comprehension is a technology that goes beyond simply understanding documents to provide optimized answers through analysis and inference, directly impacting the advancement of AI-powered real services. Kakao Enterprise will continue to strengthen research capabilities in various fields including natural language processing, vision, and speech.”

This content was produced with the assistance of AI translation services.