MARSHAL: Multiple-Attribute Regret Theory and Semantically Aware Probabilistic Weights Based Hesitant Linguistic Decision-Making

With ever-increasing abundance of text data, decision-making problems are becoming more complex. Such complexity is often a consequence of nuanced input linguistic information, which is already highly uncertain and subjective. In this article, a novel multiattribute decision-making (MADM) model named MARSHAL is introduced in order to capture the aforementioned nuanced characteristics from raw input linguistic data. It is for the first time in the literature of fuzzy based linguistic MADM models that MARSHAL treats inputs as rich linguistic features from a pretrained deep learning based large language model. This is in addition to the input being uncertain and hesitant, while also presenting vector arithmetic based information elicitation from corresponding hesitant fuzzy linguistic term sets. The idea is to introduce semantically-aware probabilistic attribute weights based on high-dimensional linguistic features learned by an enhanced variant of BERT, utilized to solve MADM problems alongside risk and regret-aversive behaviors of experts. The proposed model is tested for applicability on a real case-study of employee flight risk detection. Additionally, extensive quantitative and qualitative experiments are performed to prove the proposed model’s gained interpretability and adaptability among several other properties that existing congeneric models do not possess.