This study introduces a novel approach for computational hypothesis generation in psychology by leveraging the synergy between causal knowledge graphs and large language models (LLMs). Analyzing 43,312 psychology articles, the researchers extracted causal relation pairs using an LLM to create a specialized causal graph. Link prediction algorithms then generated 130 potential psychological hypotheses on well-being. Comparison with hypotheses from doctoral scholars and LLM-only methods revealed that the combined LLM and causal graph approach produced hypotheses comparable in novelty to expert-generated ideas, significantly outperforming LLM-only hypotheses. Deep semantic analysis corroborated these findings, demonstrating the superior conceptual incorporation and broader semantic spectrum of the combined approach. The results suggest that integrating LLMs with machine learning techniques like causal knowledge graphs can revolutionize automated discovery in psychology.

Song Tong, Kai Mao, Zhen Huang, Yukun Zhao, Kaiping Peng