Problem
A platform that matches people to jobs, courses and ads needs one shared vocabulary of skills with known relations between them. LinkedIn's taxonomy reached nearly 39k skills, 374k aliases in 26 locales and more than 200k edges, and it grew nearly 35% from February 2021 to March 2023. Taxonomists alone cannot curate this many skills as new ones keep emerging, and raw terms from profiles and search queries are noisy or ambiguous.
Example from the article: "Cadence" and "Boundary" are heavily used but unclear, so taxonomists check usage metadata and resolve them to "Cadence Software" and "Boundary Line".
Solution in brief
Skills are nodes in a directed hierarchy (parent to child) that may have several parents. Candidate terms are filtered by usage and cleaned, taxonomists place them, and a BERT-based model (KGBert) suggests parent-child relations. The model is trained on curated pairs plus four kinds of constructed negatives, and taxonomist decisions become new training data.
Questions this case answers
- How to scale the construction of a hierarchical vocabulary when manual curation cannot keep up with new terms and ambiguous names
- How to train a classifier that predicts hierarchical relations between two entities when only curated positive pairs exist and negatives must be constructed
- How to grow a hierarchy of tens of thousands of skills or tags faster than taxonomists can curate it by hand?
- How to predict parent-child relations between two taxonomy nodes with a text model?
Tools and methods
- KGBert (BERT fine-tuned on skill pairs) — classifies a skill pair as parent, child or no relation.
- Human-in-the-loop curation — taxonomists review model suggestions; their edits retrain the model.
- Constructed negative pairs — four negative types built from the curated hierarchy.
- Skill graph served via Rest.Li and HDFS — online and offline access for downstream models.
Google or a one-time email code. No password.