Poster #102 - Leveraging Large Language Model Embeddings to Predict Technology Adoption in Manufacturing

• In Event: Friday Poster Reception

Abstract

This study addresses a central challenge in understanding the diffusion of advanced manufacturing technologies: the lack of forward-looking measures capable of capturing latent adoption potential across industries and regions. Traditional indicators such as R&D intensity, patent activity, and capital investment are inherently retrospective and therefore limited in their ability to anticipate future adoption patterns. To overcome this limitation, the paper introduces a novel, task-based framework that leverages large language model (LLM) embeddings to quantify the functional alignment between technological capabilities and the tasks performed within industries.

The proposed approach integrates three primary data sources: (1) textual descriptions of Industry 4.0 technologies (e.g., artificial intelligence, robotics, cloud computing), (2) detailed occupational task data from O*NET and the Standard Occupational Classification system, and (3) employment data from the 2024 Occupational Employment and Wage Statistics (OEWS). Using advanced embedding models such as all-mpnet-base-v2 and text-embedding-3, both technology descriptions and task profiles are transformed into high-dimensional vector representations. Semantic similarity, measured via cosine similarity, captures the degree of alignment between technological capabilities and occupational tasks. These task-level relevance scores are then aggregated, weighted by employment shares, to construct industry- and region-level indices of technological relevance.

Empirically, the study evaluates this framework in two stages. First, it assesses descriptive validity by comparing the derived relevance scores with observed adoption patterns across industries and regions, using visualizations and correlation analyses. Second, a multivariate modeling framework is employed to test the predictive power of the semantic measure. Logistic regression models estimate the probability of technology adoption as a function of the relevance score, controlling for traditional determinants such as R&D intensity and intellectual property activity.

The study contributes conceptually by advancing a task-based perspective on technology adoption, emphasizing the micro-level alignment between technologies and work activities rather than relying on coarse industry classifications. Methodologically, it demonstrates the scalability and flexibility of LLM-based embeddings as a tool for measuring technological relevance, particularly in rapidly evolving technological landscapes.

From a policy standpoint, the framework provides a forward-looking indicator that can help identify regions and sectors with high latent adoption potential. This enables more targeted interventions in workforce development, training, and infrastructure investment. Overall, the study offers a novel, data-driven approach to anticipating technological change, bridging advances in natural language processing with economic analysis to improve both forecasting and policy design.

Author

• Chaneum Park, Georgia Institute of Technology
  Presenting Author