Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sensivity of LLMs' Explanations to the Training Randomness:Context, Class & Task Dependencies

Romain Loncour, Jérémie Bogaert, François-Xavier Standaert

TL;DR

This paper investigates how the (syntactic) context, the classes to be learned and the tasks influence this explanations'sensitivity to randomness, and shows that they all have statistically significant impact.

Abstract

Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks.

Sensivity of LLMs' Explanations to the Training Randomness:Context, Class & Task Dependencies

TL;DR

This paper investigates how the (syntactic) context, the classes to be learned and the tasks influence this explanations'sensitivity to randomness, and shows that they all have statistically significant impact.

Abstract

Transformer models are now a cornerstone in natural language processing. Yet, explaining their decisions remains a challenge. It was shown recently that the same model trained on the same data with a different randomness can lead to very different explanations. In this paper, we investigate how the (syntactic) context, the classes to be learned and the tasks influence this explanations' sensitivity to randomness. We show that they all have statistically significant impact: smallest for the (syntactic) context, medium for the classes and largest for the tasks.
Paper Structure (8 sections, 6 figures)

This paper contains 8 sections, 6 figures.

Table of Contents

  1. Introduction
  2. Experiment design
  3. Methods
  4. Metric
  5. Impact of the syntactic context
  6. Class dependency: absence of discriminant words
  7. Real world use-case : ArXiv vs InfOpinions
  8. Conclusion

Figures (6)

  • Figure 1: MCWME estimation with Pearson's correlation coefficient.
  • Figure 2: MCWME comparison plot between shuffled and non-shuffled sentences. A red cross on top of the plot highlights a non significant difference
  • Figure 3: Explanations' boxplot for a text containing the first name "John".
  • Figure 4: MCWME comparison plot. Texts of the left class always contain the first name "John". Texts on the right contain either "James" or no first name.
  • Figure 5: Explanations' boxplot for a text where "John" is replaced by "today".
  • ...and 1 more figures