Improving Causal Interventions in Amnesic Probing with Mean Projection or LEACE

TL;DR

This paper introduces Mean Projection and LEACE as more precise alternatives to INLP for amnesic probing, improving the removal of targeted linguistic information to better understand model behavior.

Contribution

It proposes and evaluates Mean Projection and LEACE methods, demonstrating their effectiveness in more accurately removing specific information compared to INLP.

Findings

MP and LEACE remove information more precisely than INLP

Enhanced ability to interpret model behavior through targeted information removal

Improved performance in amnesic probing tasks

Abstract

Amnesic probing is a technique used to examine the influence of specific linguistic information on the behaviour of a model. This involves identifying and removing the relevant information and then assessing whether the model's performance on the main task changes. If the removed information is relevant, the model's performance should decline. The difficulty with this approach lies in removing only the target information while leaving other information unchanged. It has been shown that Iterative Nullspace Projection (INLP), a widely used removal technique, introduces random modifications to representations when eliminating target information. We demonstrate that Mean Projection (MP) and LEACE, two proposed alternatives, remove information in a more targeted manner, thereby enhancing the potential for obtaining behavioural explanations through Amnesic Probing.