Rank-One Editing of Encoder-Decoder Models

TL;DR

This paper introduces a rank-one editing method for encoder-decoder models, enabling effective behavior deletion with minimal data, offering a direct intervention alternative to retraining or finetuning.

Contribution

It proposes a novel rank-one editing algorithm for encoder-decoder models that efficiently deletes undesired behaviors using only a single positive example.

Findings

High efficacy in deleting behaviors with minimal data

Requires only a single positive example for effective editing

Applicable to multiple NMT editing tasks

Abstract

Large sequence to sequence models for tasks such as Neural Machine Translation (NMT) are usually trained over hundreds of millions of samples. However, training is just the origin of a model's life-cycle. Real-world deployments of models require further behavioral adaptations as new requirements emerge or shortcomings become known. Typically, in the space of model behaviors, behavior deletion requests are addressed through model retrainings whereas model finetuning is done to address behavior addition requests, both procedures being instances of data-based model intervention. In this work, we present a preliminary study investigating rank-one editing as a direct intervention method for behavior deletion requests in encoder-decoder transformer models. We propose four editing tasks for NMT and show that the proposed editing algorithm achieves high efficacy, while requiring only a single instance of positive example to fix an erroneous (negative) model behavior.