Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring the Trade-off Between Model Performance and Explanation Plausibility of Text Classifiers Using Human Rationales

Lucas E. Resck, Marcos M. Raimundo, Jorge Poco

TL;DR

This work tackles the gap between faithful yet potentially implausible post-hoc explanations and human intuition in text classifiers. It introduces a model- and explainer-agnostic framework that injects human rationales during training via a contrastive-inspired loss, optimized alongside the standard cross-entropy loss in a multi-objective setup to trace a Pareto frontier between accuracy and plausibility. Empirical results across DistilBERT, BERT-Mini, and TF-IDF models on HateXplain, TSE, and Movie Reviews show improved explanation plausibility (via LIME/SHAP) with small or negligible decreases in predictive performance, and enhanced faithfulness metrics like sufficiency. The approach is robust across explainers and datasets, offering a practical path to more trustworthy NLP models while preserving flexibility and scalability for diverse architectures and tasks.

Abstract

Saliency post-hoc explainability methods are important tools for understanding increasingly complex NLP models. While these methods can reflect the model's reasoning, they may not align with human intuition, making the explanations not plausible. In this work, we present a methodology for incorporating rationales, which are text annotations explaining human decisions, into text classification models. This incorporation enhances the plausibility of post-hoc explanations while preserving their faithfulness. Our approach is agnostic to model architectures and explainability methods. We introduce the rationales during model training by augmenting the standard cross-entropy loss with a novel loss function inspired by contrastive learning. By leveraging a multi-objective optimization algorithm, we explore the trade-off between the two loss functions and generate a Pareto-optimal frontier of models that balance performance and plausibility. Through extensive experiments involving diverse models, datasets, and explainability methods, we demonstrate that our approach significantly enhances the quality of model explanations without causing substantial (sometimes negligible) degradation in the original model's performance.

Exploring the Trade-off Between Model Performance and Explanation Plausibility of Text Classifiers Using Human Rationales

TL;DR

This work tackles the gap between faithful yet potentially implausible post-hoc explanations and human intuition in text classifiers. It introduces a model- and explainer-agnostic framework that injects human rationales during training via a contrastive-inspired loss, optimized alongside the standard cross-entropy loss in a multi-objective setup to trace a Pareto frontier between accuracy and plausibility. Empirical results across DistilBERT, BERT-Mini, and TF-IDF models on HateXplain, TSE, and Movie Reviews show improved explanation plausibility (via LIME/SHAP) with small or negligible decreases in predictive performance, and enhanced faithfulness metrics like sufficiency. The approach is robust across explainers and datasets, offering a practical path to more trustworthy NLP models while preserving flexibility and scalability for diverse architectures and tasks.

Abstract

Saliency post-hoc explainability methods are important tools for understanding increasingly complex NLP models. While these methods can reflect the model's reasoning, they may not align with human intuition, making the explanations not plausible. In this work, we present a methodology for incorporating rationales, which are text annotations explaining human decisions, into text classification models. This incorporation enhances the plausibility of post-hoc explanations while preserving their faithfulness. Our approach is agnostic to model architectures and explainability methods. We introduce the rationales during model training by augmenting the standard cross-entropy loss with a novel loss function inspired by contrastive learning. By leveraging a multi-objective optimization algorithm, we explore the trade-off between the two loss functions and generate a Pareto-optimal frontier of models that balance performance and plausibility. Through extensive experiments involving diverse models, datasets, and explainability methods, we demonstrate that our approach significantly enhances the quality of model explanations without causing substantial (sometimes negligible) degradation in the original model's performance.
Paper Structure (57 sections, 2 theorems, 7 equations, 22 figures, 8 tables)

This paper contains 57 sections, 2 theorems, 7 equations, 22 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Use of Rationales.
  4. Performance and Explainability Trade-off.
  5. Theoretical Background
  6. Explainability
  7. Rationale.
  8. Explanation Plausibility.
  9. Explanation Faithfulness.
  10. Multi-objective Optimization
  11. Methodology
  12. Notation Description
  13. Contrastive Rationale Loss
  14. Trade-off Exploration
  15. Experiments
  16. ...and 42 more sections

Key Result

Theorem 1

If $w \in (\mathbb{R}_+^{*})^m$ and $x^*$ is a solution of the weighted problem, then $x^*$ is a Pareto-optimal solution of the original MOO problem.

Figures (22)

  • Figure 1: Examples of local saliency post-hoc explanations from a hypothetical text classifier for a positive movie review. Explanation (a) is more plausible than (b). Green means a positive contribution to the model's prediction, and red is negative.
  • Figure 2: Examples of explanations of the hate speech class. Explanation (a) is from the original model, and (b) is from the model with top-AUPRC. Green means a positive contribution to the model's prediction. The top-1 token was selected for visualization purposes. More examples in \ref{['tab:more_bad_explanations']}.
  • Figure 3: (a) Trade-off between the two losses on the training data. (b) Trade-off between accuracy and plausibility of the test data. The color scale represents the cross-entropy weight $w_1$ (\ref{['sec:trade_off']}). We ignore the model with $w_1 = 0$ as it is out of scale. Results including $w_1 = 0$ and shared scale between axes are in \ref{['appendix:additional_results']}.
  • Figure 4: Trade-off between accuracy and faithfulness (sufficiency and comprehensiveness) on test data. Higher values are better. The color scale is the same as the previous figures. The data scale is equal between the two graphics and their x- and y-axes.
  • Figure 5: Trade-offs between performance (accuracy, x-axis) and plausibility (AUPRC, y-axis, in percentage (%)) for all models and datasets (test data). There are 2 random (negative) rationales, and the explainer is LIME. Green dots are the models chosen to be analyzed more carefully. The color scale is the same as the previous figures. We ignore the model with $w_1 = 0$ in all graphics as it is out of scale. Larger figure and results including $w_1 = 0$, 5 rationales and/or SHAP, shared scale between axes, and Pareto-frontiers are in \ref{['appendix:additional_results']}.
  • ...and 17 more figures

Theorems & Definitions (7)

  • Definition 3.1: Multi-objective optimization problem
  • Definition 3.2: Pareto-optimality
  • Definition A.1: Weighted sum method
  • Theorem 1: Necessity
  • proof
  • Theorem 2: Sufficiency
  • proof