KETI Unveils 'HAECHI,' Korean Culture Heritage and Hangul-Specialized AI Model

This page was machine-translated and may differ from the original. View original

KETI Unveils 'HAECHI,' Korean Culture Heritage and Hangul-Specialized AI Model

Published2026.10.01 09:11

Culture Heritage Identification Accuracy Enhanced by Up to 74 Percentage Points, Model Released on Hugging Face

The Korea Electronics Technology Institute (KETI) has developed two versions of the 'HAECHI' AI model that identifies Hangul and Korean cultural heritage in images and processes complex web and app-based tasks. To address the limitation that foreign language models fail to properly recognize Korean cultural heritage and Hangul, the institute adopted an approach of adding self-constructed training data to existing models for additional training.

The Natural Language Processing Research Team at KETI's AI-Based Model Research Center announced on the 1st that it has constructed two versions of HAECHI by providing additional training to models based on HyperCLOVA X and Qwen3.5 using training data including Korean objects and cultural heritage images, Hangul signage, and public documents.

HAECHI simultaneously processes images and text, and can also perform multi-stage tasks such as ticket issuance, product purchase, refunds, and service subscription.

As a result of measuring performance using the H400 evaluation set (targeting 400 Korean cultural heritage items) constructed by the research team, the accuracy rate in correctly identifying cultural heritage names improved from 0.92% to 38.07% for the Qwen3.5-based model and from 5.05% to 74.16% for the HyperCLOVA X-based model, respectively.

Hangul signage recognition ability in street photos also improved by more than 5 percentage points for both models.

The Qwen3.5-based model increased from 36.13% to 41.41%, and the HyperCLOVA X-based model increased from 33.59% to 38.87%.

In the evaluation of long-task execution capability (Tau2), which determines the next action based on the results of the preceding task, comprehensive scores for both models rose by approximately 2 percentage points.

This model development received GPU resource support through the National IT Industry Promotion Agency (NIPA)'s 'Advanced GPU Utilization Support Program,' and the Korean cultural heritage dataset was constructed through a project under the Korea Information Technology Planning and Evaluation Institute (IITP)'s 'Human-Centered Artificial Intelligence Core Technology Development Program.'

Kim Jung-ho, Director of KETI's AI-Based Model Research Center, stated, "Through AI that understands Korean culture and language well, we will be able to properly introduce Korea to the world," and added, "We will contribute to establishing an independent foundation model ecosystem with the Korea-specialized data built this time and develop HAECHI into Korean-style AI agent technology."

The model files, evaluation results, and execution methods for the two versions of HAECHI can be found on Hugging Face.

#KETI #AI 모델 #멀티모달 #Qwen #자연어처리 #HAECHI #한글 인식 #문화유산 식별 #HyperCLOVA X

To request a correction, reply or follow-up report on this article, see how to file a request. Previously published statements are collected in corrections & replies.