Scholarlink

This paper presents a pipeline for extracting, classifying, and representing skills requested in job advertisements to enable demand-side labor market analysis. Our contributions include: (1) addressing the annotation bottleneck by leveraging scalable, taxonomy-aligned LLM supervision for training a lightweight sentence encoder, (2) expanding skill extraction to include implicit skill requirements as well as the explicit mentions typically targeted in prior work, and (3) representing skills as distributions to robustly support downstream tasks despite the fluid, overlapping nature of skill definitions. Concretely, we compile 3M+ postings from 10k+ sources and sample 500k+ sentences to fine-tune paraphrase-multilingual-mpnet-base-v2 for identifying skill requests and mapping them to the 13,896-skill ESCO taxonomy, supervised by GPT-4o mini. The outcome is normalized per-ad skill distributions, aggregated from sentence-level distributions weighted by request probability.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-Supervised Multilingual Skill Extraction and Classification from Job Ads

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jakob Mørup Wang,
  • Zhiru Sun

摘要

This paper presents a pipeline for extracting, classifying, and representing skills requested in job advertisements to enable demand-side labor market analysis. Our contributions include: (1) addressing the annotation bottleneck by leveraging scalable, taxonomy-aligned LLM supervision for training a lightweight sentence encoder, (2) expanding skill extraction to include implicit skill requirements as well as the explicit mentions typically targeted in prior work, and (3) representing skills as distributions to robustly support downstream tasks despite the fluid, overlapping nature of skill definitions. Concretely, we compile 3M+ postings from 10k+ sources and sample 500k+ sentences to fine-tune paraphrase-multilingual-mpnet-base-v2 for identifying skill requests and mapping them to the 13,896-skill ESCO taxonomy, supervised by GPT-4o mini. The outcome is normalized per-ad skill distributions, aggregated from sentence-level distributions weighted by request probability.