Enhancing Group Decision-Making through Large Language Models for Semantic Filtering of Expert Opinions

Abstract

In modern decision-making contexts involving multiple experts, the use of natural language to express opinions introduces considerable difficulties for traditional analytical methods. With the growth of online collaboration, expert evaluations have become increasingly diverse in both format and scale, especially when communicated through unstructured textual comments. While conventional Group Decision-Making techniques can handle varied evaluation metrics, they cannot often fully interpret the semantic richness of human language. This work presents a novel Group Decision-Making framework that leverages a Large Language Model to semantically analyse and filter expert commentary, specifically DeepSeek LLM, identifying and removing remarks that may negatively influence others. Unlike standard sentiment analysis approaches based on fixed lexical rules, the Large Language Model captures subtle linguistic features such as tone, context, and implied meaning, allowing for a more accurate extraction of genuine expert preferences. The filtered and interpreted inputs are then aggregated to produce a consensus that better reflects the true collective judgment. This integration of advanced natural language processing enhances the interpretability, fairness, and reliability of group decisions in complex and dynamic environments. This constitutes a departure from traditional sentiment analysis approaches, as the LLM is used not only for interpretation but also for active semantic filtering of expert discourse.