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

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.

A Question on the Explainability of Large Language Models and the Word-Level Univariate First-Order Plausibility Assumption

The explanations of large language models have recently been shown to be sensitive to the randomness used for their training, creating a need to characterize this sensitivity. In this paper, we propose a characterization that questions the possibility to provide simple and informative explanations for such models. To this end, we give statistical definitions for the explanations' signal, noise and signal-to-noise ratio. We highlight that, in a typical case study where word-level univariate explanations are analyzed with first-order statistical tools, the explanations of simple feature-based models carry more signal and less noise than those of transformer ones. We then discuss the possibility to improve these results with alternative definitions of signal and noise that would capture more complex explanations and analysis methods, while also questioning the tradeoff with their plausibility for readers.