Background

Knowledge editing is the task of modifying the knowledge a model possesses with minimal intervention.

Problem Definition

• Korean laws and regulations change frequently, so there is a need for continuous knowledge editing.
• The meaning of laws can change significantly with minor differences in articles, clauses, and items, so it is necessary to effectively and fine-grainedly change only the desired statutes.

Solution

We constructed a knowledge book to enable real-time knowledge editing and trained a retriever to fetch information from it effectively. Additionally, we separately trained an LLM to effectively incorporate the retrieved knowledge.

Achievements

• Achieved a performance improvement of about 15% in knowledge editing compared to existing methods.
• Accepted to EMNLP 2025 Findings.

My Role

• As a co-author, I implemented and experimented with baseline methods for knowledge editing.
• I proposed and implemented a hard negative data mining method by subtly changing legal content by altering articles, clauses, and items one by one.
• I trained the BGE-m3 and E5 embedding models with the constructed hard negative samples using the BGE-m3 loss function.