A System for Recognizing Required Skills in Job Postings: Skill Profiling

Alexander Kovalev, Kirill Usenko25 May 2024

Problem

A customer wants to know which skills are in demand on the labor market and which pay the most. The raw material is job postings from several aggregators. The obvious shortcut is to use the employer's "key skills" field as the target. About 50% of postings have no such field, and requiring enough skills per posting leaves only 10–15% of the data. The listed skills are also too generic ("1C" can mean accounting or trade management, priced differently) and miss many duties.

Example from the article: in a gas-equipment fitter posting, the annotation held one skill, "performing fitter work", while the model also found duties such as disconnecting consumers from the gas distribution network.

Solution in brief

Annotate skills and duties directly in the posting text with BIO tags and fine-tune a BERT-type encoder as a token classifier (NER). Evaluate with span-level F1, send disagreements back to the customer for relabeling, and repair broken tags with two rules. Test result: F1-Strict 0.60, F1-Loose macro 0.83.

Questions this case answers
  • How to extract skill and duty mentions from free-text postings as labeled spans when the structured field is missing or too generic, and evaluate span extraction fairly
  • How to repair inconsistent token-level tags from a sequence labeler so that the resulting spans are valid
  • How to extract required skills from job postings when the employer's key-skills field is missing or too generic
  • How to label and train a NER model that finds multi-word spans such as skills and duties

Tools and methods

  • NER with BIO tags — marks each token as skill start (B), continuation (I) or other (O).
  • DeepPavlov rubert, XLM-RoBERTa, T5 encoder, SBER rubert — encoders compared.
  • F1-Strict, F1-Loose macro, F1-Loose micro — span-level metrics.
  • Post-processing heuristics — fix I tags that have no preceding B.
  • ChatGPT API, embedding clustering — tried and rejected.
Continue reading — it's free

Google or a one-time email code. No password.