dnotitia.com, Aug. 28, 2025 –
[Seoul, South Korea – August 28th, 2025] Dnotitia Inc. (Dnotitia), a specialized company in long-term memory AI and semiconductor-integrated solutions, announced that the research project it participated in, focused on building a multimodal dataset reflecting Korean cultural heritage, has been accepted to the Dataset Track of ACM Multimedia 2025, one of the world’s most authoritative conferences in the multimedia field.
Since its inception in 1993, ACM Multimedia has grown into one of the world’s premier conferences in multimedia, covering multimodal AI and next-generation media technologies. The conference addresses the full spectrum of research across image, video, speech, and text processing and integration. Each year, thousands of papers are submitted, but only a small fraction are accepted, reflecting the event’s highly competitive nature. The 33rd edition will be take place from October 27 – 31, 2025. in Dublin, Ireland.
The accepted paper introduces HAN (Heritage Augmented Narrative Visual-Language Description Dataset), a multimodal dataset designed to reflect Korea’s cultural heritage and linguistic nuances. Going beyond simple translation, HAN adopts narrative-style captions that capture emotional context, social interactions, and cultural storytelling, helping AI systems overcome performance limitations of existing multimodal models and generalize across diverse, multicultural environments
The HAN dataset goes beyond simple translation, adopting narrative-style captions that capture emotional context, social interactions, cultural background, and storytelling. This approach helps overcome the performance limitations of existing multimodal models, while alleviating bias issues common in large-scale image-text training and supporting the development of vision-language models that can generalize across multicultural environments.
The study is characterized by the systematic construction of a dataset extracted from 7,822 Korean broadcast programs, including 41,000 images and 410,000 Korean-English narrative captions. A key achievement is overcoming the limitations of English-centric datasets, thereby correcting linguistic imbalances and cultural bias while making multilingual and multicultural AI learning possible, including for low-resource languages such as Korean. In addition, by offering a scalable alternative to the traditionally high-cost, expert-dependent process of building cultural and language-based datasets, the project represents a significant breakthrough for future research.
In particular, by applying a narrative-style captioning approach, HAN richly captures the contextual meaning of cultural heritage, an innovation that has drawn attention from both academia and industry for combining innovative ideas with demonstrated practical value.
Furthermore, to validate the dataset’s effectiveness, the research team conducted follow-up experiments using the diversity of narrative captions, which resulted in significant performance improvements over existing models. This confirmed that the HAN dataset is not just a data collection effort but a resource with practical value for both academic research and real-world applications – providing strong evidence of its broader impact.
As a foundational dataset, HAN is expected to contribute significantly to the global AI research ecosystem, with potential applications across multimodal AI, natural language processing, and digital archiving of cultural heritage.
“Just as K-pop and K-dramas have become part of daily life around the world, the time has come for AI models to also embody Korean culture,” said Moo-Kyoung Chung, CEO of Dnotitia. “The HAN dataset is more than a research outcome, it represents a first step in allowing Korean culture to permeate global AI models, and will play a key role in ensuring data diversity and reducing bias across the AI ecosystem.”
This achievement is part of the “Korean Cultural Video Understanding Dataset” project, supported by Korea’s Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA). Building on this milestone, Dnotitia plans to further expand its efforts by developing multimodal AI datasets that reflect Korean cultural heritage and linguistic diversity. They company aims not only to drive technical innovation but also to foster a more inclusive and equitable AI ecosystem.
[About Dnotitia]
Dnotitia is an AI and semiconductor company that creates innovative value through the convergence of artificial intelligence (AI) and data, providing high-performance and low-cost LLM solutions. Leveraging Dnotitia’s world’s first Vector Data Processing Unit (VDPU), the company offers ▲Seahorse, the high-performance vector database, which supports Retrieval-Augmented Generation (RAG) solutions – a key technology for Gen AI. Additionally, Dnotitia offers ▲Mnemos, a personal/Edge LLM device based on its proprietary LLM foundation model.
① Seahorse is indexes various types of multi-modal data, such as text, images, and videos, into vector form, providing semantic search that extracts information reflecting meanings and contexts based on user queries. Seahorse can be used not only in RAG systems but also for implementing semantic search across all digital data stored globally.
② Mnemos is a solution designed to address the high costs and resource consumption of AI. It is a compact edge device capable of running high-performance LLM without the need for a data center. Leveraging Dnotitia’s RAG and LLM optimization technology, Mnemos delivers high-performance LLM services using minimal GPU/NPU resources.
Founded in 2023, Dnotitia has grown to a team of over 100 employees in a short period of time and has established strategic partnerships across various industries. By integrating specialized semiconductors and optimized algorithms, Dnotitia aims to usher in a new era of AI. Through the fusion of data with AI to develop AI with long-term memory, Dnotitia envisions realizing a low-cost AGI (Artificial General Intelligence) accessible to everyone, creating a future where the benefits of AI can be enjoyed by all.
For more information about Dnotitia, please visit: www.dnotitia.com .