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

TL;DR

This paper introduces MEMAT, a novel method for cross-lingual knowledge editing in transformers that leverages attention mechanisms to improve factual accuracy with minimal parameter changes.

Contribution

The paper proposes MEMAT, a new approach that enhances knowledge editing across languages by focusing on attention mechanisms, outperforming existing methods.

Findings

10% increase in magnitude metrics

Effective across multiple languages including unseen ones

High portability of the editing method

Abstract

Recent research has explored methods for updating and modifying factual knowledge in large language models, often focusing on specific multi-layer perceptron blocks. This study expands on this work by examining the effectiveness of existing knowledge editing methods across languages and delving into the role of attention mechanisms in this process. Drawing from the insights gained, we propose Mass-Editing Memory with Attention in Transformers (MEMAT), a method that achieves significant improvements in all metrics while requiring minimal parameter modifications. MEMAT delivers a remarkable 10% increase in magnitude metrics, benefits languages not included in the training data and also demonstrates a high degree of portability. Our code and data are at https://github.com/dtamayo-nlp/MEMAT.