Identification of Responsibility-Skill Coupling Patterns in Embodied Intelligent Manufacturing Driven by Improved BERTopic and Association Rules

INTRODUCTION: Embodied Intelligence (EI) is rapidly penetrating industrial manufacturing, yet the interdisciplinary nature of EI systems makes it challenging for enterprises to precisely identify evolving talent demands. OBJECTIVES: This study aims to identify responsibility topics and responsibility–skill coupling patterns in embodied intelligent manufacturing, thereby providing a theoretical basis for talent demand analysis. METHODS: A total of 749 job advertisements were scraped and an improved BERTopic method integrating GSDMM topic similarity with semantic representation was proposed, identifying 14 responsibility topics (10 domain-specific)  across four clusters. The Apriori algorithm was then applied to mine association rules between responsibility topics and skill keywords, with thresholds determined by the elbow method. RESULTS: GSDMM-BERTopic consistently outperformed classic BERTopic and Gibbs-BERTopic, achieving average CV coherence of 0.5477, silhouette score of 0.6737, and outlier ratio of 0.1662. The selected configuration yielded 14 responsibility topics (10 domain-specific) across four clusters. Association rule mining produced 128 filtered rules. Notable topic–skill couplings include 3D Perception Modeling with Detection and Segmentation (lift 9.015), Motion Control with Reinforcement Learning and ROS (lift 2.277), and Frontier Technology Tracking with Ph.D. (lift 2.605). Cross-category analysis reveals Python and PyTorch as the common skill substrate, C++ as critical in hardware-adjacent domains, and the ROS/ROS2 ecosystem as the central middleware. CONCLUSION: The findings reveal a T-shaped skill profile in embodied intelligent manufacturing and provide practical guidance for enterprises to optimize workforce allocation and talent cultivation strategies.