Add new paper:

[wyzh0912]

Title

Mass-Editing Memory with Attention in Transformers: A cross-lingual exploration of knowledge

Published Date

2024-08-01

Source

ACL

Head Name

Memory Head

Summary

• Innovation: The paper introduces Mass-Editing Memory with Attention in Transformers (MEMAT), a method that significantly improves cross-lingual knowledge editing in transformer models by minimally modifying parameters and optimizing specific attention heads.

• Tasks: The study uses cross-lingual datasets in English and Catalan to evaluate the effectiveness of MEMAT, focusing on factual knowledge editing in language models through experiments that assess the efficacy, generalization, and specificity of knowledge edits using various prompts and interventions.

• Significant Result: MEMAT achieves more than a 10% improvement in magnitude metrics compared to previous methods, demonstrating enhanced portability and computational efficiency in cross-lingual knowledge editing without significant degradation in existing knowledge.