New language model for advanced Korean-based AI service developed
12 June 2019

New language model for advanced Korean-based AI service developed

SEOUL, June 11 (Yonhap) -- A new Korean language model that can advance artificial intelligence (AI) services and searches has been developed for a wide range of applications, a state-run electronics institute said Tuesday. The system has an average 4.5 percent better processing and understanding capability compared with Google Inc.'s widely used Bidirectional Encoder Representations from Transformers (BERT) deep language learning tool, according to the Electronics and Telecommunications Research Institute (ETRI). The institute said while the new KorBERT is based on BERT that converts words into numbers, it has made full use of the extensive data collected in the last 10 years from local Korean language news articles and encyclopedias. "The new model, which is part of South Korea's larger Exo-brain AI project, takes into account the Korean language's distinct characteristics like postpositional particles coming after a root word, that helps better understanding and faster processing," ETRI said. The institute claims this can lay the foundation for creating a viable AI secretarial service that can better handle questions, conduct searches for materials requested by users and overcome some of the limitations found in the current technology. Existing AI has drawn flak for not really being good at responding "naturally" or quickly to questions posed by people and often not understanding the questions raised at all. ETRI said it has filed some 70 patents so far related to the Exo-brain endeavor, and the new language model is fully compatible with leading deep learning frameworks, such as PyTorch and Tensorflow, and is offered as an open application programming interface so people and companies can access it to meet their needs. It said the model has most recently been used in the early beta version of Hancom Office's search engine, with the system to be employed in the question and answer search tool for legal matters in the second half of 2019. It will, moreover, be developed for use to intelligently search for details on patents, ETRI said. [email protected] (END)