Techday
Physical AI HBM Smart Factory SDV AIoT Power Semicon 특수 가스 정정·반론보도 모음 e4ds plus

SKT-National Institute of the Korean Language Cooperate to Develop ‘Human-Level’ AI Korean Language Model

Google 우선 소스 기사입력2021.04.08 08:15

SKT, National Institute of the Korean Language Develop Korean Language Model
Expected huge GLM with 150 billion parameters
Increased accuracy expected to be utilized in various industrial fields



On the 7th, SK Telecom signed a business agreement with the National Institute of the Korean Language and decided to develop a next-generation AI language model optimized for the Korean language using the National Institute of the Korean Language's language information.
▲ (From left) SKT Davis Eric Hartman, Research Director
National Institute of the Korean Language Jeong Hee-won, Director of the Language Research Lab [Photo = SKT]

The next-generation AI Korean model is a Korean general-purpose language model (GLM) with similar performance to the English-based GPT-3 language model developed by the OpenAI Research Institute. It can implement human-level GPT-3 functions in Korean, such as solving language-related problems, writing, translation, and performing simple coding based on given sentences.

GLM can be applied to a wide range of language activities, including everyday emotional conversations, customer center conversations in various industries, as well as current affairs, literature, history, and games, and it is expected that it will be further utilized in new industries based on this.

The GLM being developed by SKT this time is a huge language model with 150 billion parameters. Considering that GPT-3 has 175 billion parameters, which is 100 times larger than its predecessor, GPT-2, and has much higher accuracy and wider applicability, GLM is expected to be the definitive version of the Korean AI language model.

SKT plans to develop GLM by the end of this year, verify the model performance through internal services, and then commercialize it, and apply it to various services. It also plans to develop a Korean language model performance evaluation method and conduct research on Korean data quality evaluation.

SKT, which has been developing an AI language model since 2018, developed KoBERT in 2019 and is using it for chatbots, etc. In April 2020, it completed the development of KoGPT-2 to develop chatbot conversations more naturally, and in October 2020, it developed KoBART with improved text processing performance, strengthening its technological capabilities in the field of natural language understanding and processing.

The National Institute of the Korean Language will hold the '2021 Korean Language Information Processing System Competition' during Hangeul Week, reorganizing it to compete in AI's language communication skills using SKT's AI language model.
이수민 기자
기사 전체보기