NC and Korea University Release AI Conversation Dataset 'FoCus Dataset'

[Asia Economy Reporter Kang Nahum] NCSoft announced on the 14th the release of the artificial intelligence (AI) conversation dataset ‘FoCus Dataset (For Customized conversation dataset)’, developed through joint research with Professor Lim Heeseok’s research team at Korea University.

The FoCus Dataset is the world’s first AI conversation dataset that utilizes both the user’s personal persona and external knowledge. It consists of more than 15,000 conversations covering approximately 8,000 broad topics.

The AI applied with this dataset can understand the user’s experience, preferences, possessions, and interests during the conversation, and acquire the latest knowledge on the topic from Wikipedia in real time, enabling natural conversations.

The FoCus Dataset is characterized by enabling the implementation of conversational technology with the same performance without using ultra-large language models. Currently, general large-scale language models require significant costs for training and inference, yet show limitations in reflecting real-time knowledge and personal experiences.

The joint research team published and presented the related research paper at the world’s most prestigious AI conference, ‘AAAI 2022’, last February. In October, they will co-host a workshop (The 1st Workshop on Customized Chat Grounding Persona and Knowledge) with Korea University to share research achievements, including a data utilization competition (Shared Tasks), at the world computational linguistics conference ‘COLING 2022’.

Lee Yeonsu, head of NC Language AI Lab, said, “Recently, in the NLP academic community, new conversational technologies comparable to those based on ultra-large language models are being proposed due to cost and environmental issues. We decided to release this dataset in agreement with this research direction and hope that active discussions and technological developments will take place in the global research community.”