Interpretation of NLP models through input marginalization

TL;DR

This paper addresses the out-of-distribution issue in NLP model interpretation methods that erase tokens by proposing a marginalization approach, leading to more accurate explanations of model predictions.

Contribution

It introduces a novel token marginalization technique to improve the reliability of NLP model interpretations over existing erasure-based methods.

Findings

The proposed method provides more faithful interpretations.

It effectively addresses the out-of-distribution problem.

The approach is applicable to sentiment analysis and natural language inference models.

Abstract

To demystify the "black box" property of deep neural networks for natural language processing (NLP), several methods have been proposed to interpret their predictions by measuring the change in prediction probability after erasing each token of an input. Since existing methods replace each token with a predefined value (i.e., zero), the resulting sentence lies out of the training data distribution, yielding misleading interpretations. In this study, we raise the out-of-distribution problem induced by the existing interpretation methods and present a remedy; we propose to marginalize each token out. We interpret various NLP models trained for sentiment analysis and natural language inference using the proposed method.