NKIS 국가정책연구포털

국문초록

영문초록

• Abstract
  
  Finding suitable substitutes in a specific context is a core challenge for recommendation and information retrieval systems. In the labor market, identifying appropriate skill substitutes can alleviate skill mismatch—the discrepancy between employer demand and workforce supply. Yet, existing context-agnostic methods fail to capture functional interchangeability. We propose a framework using a Masked Language Model (MLM) trained on 3.7 billion job postings (2010–2025) to extract contextual substitutes. Our model significantly outperforms word2vec, conditional probability, and GPT-5.1 baselines, revealing that substitutability varies dynamically across occupations and time. Mechanistic analysis using Sparse Autoencoders (SAEs) demonstrates that deeper layers progressively disentangle context-specific semantics to ground these substitutions. Finally, simulations show that replacing required skills with trending, contextually compatible alternatives can increase job posting visibility by up to three orders of magnitude, offering a scalable solution for navigating the evolving labor market.

목차

• 1. Introduction
  
  2. Methods
   2.1 Job Postings Dataset
   2.2 Preprocessing
   2.3 Pre-Training and Fine-Tuning BERT
   2.4 Interpreting Contextual Substitution via Sparse Autoencoders
  
  3. Model Validation
   3.1 Model Accuracy
   3.2 Contextual Substitution Validation
    3.2.1 Co-occurrence-based Substitutability
    3.2.2 Embedding-based Substitutability
    3.2.3 Contextual Substitutability (Our Approach)
  
  4. Findings
   4.1 Substitution by SOC Major Group
   4.2 Year-wise Substitution
   4.3 Interpreting Skill Substitution via SAEs
   4.4 Quantifying Market Exposure Gain from Contextual Substitution
  
  5. Conclusion
  Limitations
  References
  Appendix