MediaZen Participates in DeepModal and HCLT?KACL... "Paper Presented on Using LLM Technology for Unfair Contract Terms Review"

Mediazen announced on the 14th that it participated in the ‘Hangeul and Korean Language Information Processing?Korean Corpus Linguistics Society Joint Conference (HCLT?KACL)’ held over two days starting from the 11th of this month.

HCLT?KACL is an academic conference first held in October 1989 and takes place annually around Hangeul Day. This year, Naver hosted the event, and the Korea Information Science Society organized it.

Mediazen presented two papers related to large language models (LLM) jointly researched with DeepModal, a startup from the Electronics and Telecommunications Research Institute (ETRI). The papers covered △confidence measurement based on generation probability (unfair contract terms review) △fine-tuning and decoding methods for high-quality document generation. The presentations were delivered by Choi Jung-yoon, a researcher at Mediazen’s Artificial Intelligence (AI) Research Institute.

The first study is a preliminary research on whether the AI model to be applied in the ‘Inter-Ministerial Collaboration-Based AI Diffusion Project (AI Convergence Contract Review Platform Construction)’ led by the Ministry of Science and ICT and the National IT Industry Promotion Agency and promoted by the Fair Trade Commission can be applied in practice. The study researched a system that uses LLM to measure document reliability in a new way and automatically classify documents subject to review on the contract review platform. The research utilized publicly available contract review data from the Fair Trade Commission.

The company explained that the system has high applicability for contract review in sensitive fields such as law through more transparent reliability calculation than existing methods. It also added that high-reliability classification is possible even with a small amount of domain data, enabling the construction of a robust contract review system.

Researcher Choi Jung-yoon said, “The system developed through this research accurately identifies documents unrelated to contract terms among those attached by complainants,” and added, “By detecting inappropriate parts in unrelated complaints and guiding complainants, it will greatly improve the efficiency of complaint handling within the Fair Trade Commission.”

In the subsequent presentation on ‘Fine-tuning and Decoding Method Research,’ a method was proposed to automatically generate diagnoses and opinions when writing medical imaging reports. Considering the difficulty of securing large-scale data in the medical field, the study explored and evaluated the optimal method to generate high-quality documents with limited data.

The company stated, “Through experiments adjusting decoding parameters such as low-rank adaptation (LoRA) training method, temperature, and beam search, which determine creativity and diversity, we explored the optimal environment and improved model performance,” emphasizing, “This research suggests the possibility of building such a system in data-limited medical environments and will serve as an important foundation for medical AI research and experimentation.”

This content was produced with the assistance of AI translation services.