Explanation sensitivity to the randomness of large language models: the case of journalistic text classification

TL;DR

This paper investigates how randomness in training large language models affects the consistency of their explanations, revealing variability in interpretability despite stable accuracy, and proposes a simpler, more explainable model with potential enhancements.

Contribution

It demonstrates that training randomness impacts explanation stability in LLMs and introduces a simpler, more explainable model that can be improved with explanation-derived features.

Findings

Training with different seeds yields similar accuracy but variable explanations.

A simpler model offers stable explanations but lower accuracy.

Inserting explanation-derived features improves the simpler model.

Abstract

Large language models (LLMs) perform very well in several natural language processing tasks but raise explainability challenges. In this paper, we examine the effect of random elements in the training of LLMs on the explainability of their predictions. We do so on a task of opinionated journalistic text classification in French. Using a fine-tuned CamemBERT model and an explanation method based on relevance propagation, we find that training with different random seeds produces models with similar accuracy but variable explanations. We therefore claim that characterizing the explanations' statistical distribution is needed for the explainability of LLMs. We then explore a simpler model based on textual features which offers stable explanations but is less accurate. Hence, this simpler model corresponds to a different tradeoff between accuracy and explainability. We show that it can be improved by inserting features derived from CamemBERT's explanations. We finally discuss new research directions suggested by our results, in particular regarding the origin of the sensitivity observed in the training randomness.