TForMIX: A Method That Combines LLM and Multidimensional Modeling for Technological Foresight

Abstract

Technical documents, such as scientific papers and patents, are widely used as a basis for Technological Foresight (TF) processes. Typically, these analyses require identifying elements (e.g., terms) in the textual contents of these documents, which are relevant to the scientific-technological domain under investigation. Information Extraction (IE) and Natural Language Processing (NLP) techniques are useful tools to automate the identification of these elements, which is essential in TF processes that usually involve the analysis of a corpus of hundreds (and sometimes thousands) of documents. An analytical view over this corpus, based on the occurrence of those relevant elements, helps prioritize document analysis and, consequently, accelerates the whole TF process. However, building a system that provides such analytical insight is expensive. Moreover, for each domain-specific TF process, a new system would have to be built. Thus, there is a need for viable solutions to analytically explore a corpus, according to the specific requirements of each domain. This work presents Technological Foresight with Multidimensional Information eXtraction (TForMIX), a novel method for building Decision Support Systems (DSSs) that applies Named Entity Recognition (NER) and Relation Extraction (RE) while allowing multidimensional analytical exploration of entities and relations together with bibliometric data from documents. TForMIX is a flexible method that can be applied to different domains, and speeds up building DSSs for each domain. Additionally, we evaluate the applicability of the produced DSSs in TF processes by conducting a practical experiment that demonstrates that applying the method to generate DSSs, supported by IE techniques, can significantly contribute to the conduction of TF analyses. The combination of the used theories, innovative methods, and proposed practical validation highlighted the high-quality nature of the analysis in this study while offering the potential for valuable insights and contributions to the TF process.