The Experimental Evaluation of Different Explainable AI Techniques for Large Language Models

Large Language Models have in recent times been a leading technology regarding the field of Natural Language Processing. Various research has been done to further develop the capabilities of Large Language Models, enhance their performance, optimize various aspects of computation and much more. But, to fully understand and trust the predictions provided by these models, a detailed interpretability or explainability analysis needs to be implemented. To do such a thing, the paradigms of Explainable artificial intelligence are extended to the field of Large Language Models and various tools such as LIME and SHAP can be taken into consideration. The experimental evaluation of different explainable AI techniques for LLMs proposed in this paper is based on combining and specially scaling the outputs obtained from both LIME and SHAP. The analysis is then done regarding every token, so the proposed method has been named as Token-wise interpretability.