Générer une explication textuelle à l’ère des LLMs : premières expérimentations

In some applications, it is necessary to provide the user with a textual explanation with the value predicted by a fuzzy inference system. Research has shown that these explanations increase the user’s confidence in the decision and in the system itself. Our previous work demonstrated a simple way to reduce the trace to the elements that should be contained in the explanation, in the simple case of a fuzzy classifier. We used NLG (Natural Language Generation) and a surface generator to produce text. In the era of LLMs, we are investigating the possibility of modernizing this approach to determine the advantages and disadvantages of both text generation approaches. In this article, we describe the results of our initial experiments.