In the natural language processing field, foundation models are mainly developed in English-speaking countries. LINE is working on building a Japanese foundation model in collaboration with NAVER. In this session, we will demonstrate the current state of HyperCLOVA, which contains the foundation model, and show some case studies.
On LINE DEVELOPER DAY 2021, we introduced the 39B parameter model, but this year we will show the improvements and issues of the scaled-up 82B model.
We’ll be covering two use cases of HyperCLOVA: a dialog system that won 1st prize at the Dialog System Live Competition and a text-writing assistant for creating emails and reports faster.
In addition, we demonstrate evaluation results of parameter-saving models, an AI filter to detect fairness and ethics in automatically-generated text, and a robust fact-checking function.