Using large language models for building fuzzy cognitive maps

Anna Kovalenko, Alina Petukhova, Chubyr Natalia Olegovna, Veronika Rastorgueva

Keywords: fuzzy cognitive maps, large language models, fuzzy systems

TL;DR: In our research, we assess nine models across three fuzzy systems and draw conclusions regarding the reasoning abilities of LLMs in this context.

Abstract: Fuzzy Cognitive Maps (FCMs) are widely used in the modelling of complex systems, reflecting the causal relationships between key concepts. The introduction of large language models (LLMs) presents new opportunities for constructing FCMs by using large text corpora and the deep learning capabilities built into these models. This study investigates the potential of leveraging LLMs, specifically OpenAI’s GPT-4, to automate the construction of FCMs. The research assesses the efficacy of LLMs in identifying causal links through structured prompting techniques, including "chain of thought" reasoning and self-consistency-based sampling, across three domains: retail system, urban development, and brain tumor characterization. The experimental results demonstrate that while LLMs effectively extract a range of relationships, challenges remain in terms of missing, inaccurate, and overestimated causal links. These findings underscore the importance of integrating expert validation and domain-specific data augmentation to enhance the accuracy and interpretability of the model. This study makes a valuable contribution to the field of automated decision support systems, advancing AI-driven approaches for the construction of FCMs.

Submission Number: 22