A GPT-3-like Korean language model trained Naver
South Korea’s Naver Corp. has unveiled its supersized artificial intelligence (AI) platform “HyperCLOVA”, a new Korean-based language model system enabling human brain-like linguistic capacity.
During an online conference on Tuesday, the leading internet portal company claimed HyperCLOVA is the world’s top-level AI technology equipped with over 204 billion parameters, way above 175 billion parameters of currently the most outstanding AI language model GPT-3 often referred to as a global standard. The more parameters a model has, the better it can distinguish the nuances of speech, tones and even dialects.
Naver also said HyperCLOVA is the world’s first hyper-mega Korean-based language model system, different from other AI models based on English. The new system learned from more Korean data than U.S. model GPT-3 by 6,500 times. Among the learned data, Korean language accounts for 97 percent.
When HyperCLOVA is used, various AI systems can be developed to process human conversation, translation, sentence generation, summarization, and machine reading, opening new service and business opportunities, according to the company.
Naver said it plans to develop HyperCLOVA as a multimodal AI that can understand even videos and images in the future.