EFFICIENT AI-POWERED DECISION-MAKING IN SYSTEMATIC LITERATURE REVIEWS USING MULTI-EMBEDDING MODELS

Purpose: This paper presents a new Artificial Intelligence (AI) multi-model approach for selecting publications that fit researcher needs well. It is the second work in a series of AI-powered systematic literature reviews focusing on replacing the human research team with a multi-model framework. Need for the study: The expected exponential growth of scientific literature generated fully or partially by AI demands advanced artificial intelligence tools to streamline systematic literature reviews. Such a method should filter databases according to the researcher's needs, minimizing the risk of overlooking critical works. Methodology: The technology behind word embedding models (WEMs) powers various AI systems, including ChatGPT. Eight models of this kind were evaluated with bibliometric criteria to select and utilize the four top-performing ones in article selection. Each model independently selects publications that best fit the problem description. Papers selected by the larger number of models are considered more relevant. Findings: Different WEMs are trained on different data, have different architectures, and recognize different publications as the most relevant. Usually, articles selected by more models show higher thematic coherence; however, the number of models in the framework cannot be too large. The process is analogous to a systematic literature review conducted by many researchers with all the advantages and disadvantages of their parallel work. Practical Implications: Our method is less biased and more accurate than a multiple-person systematic literature review (SLR). It allows for a fast examination of the problem from many perspectives and has the same implications as SLR.