Skill Graph Construction From Semantic Understanding
Shiyong Lin, Yiping Yuan, Carol Jin, Yi Pan · 2023
LinkedIn is building a skill graph to power a skill-first talent marketplace. Constructing a skill graph from a flat list is not an trivial task, especially by human curation. In this paper, we leverage the pre-trained large language model BERT to achieve this through semantic understanding on synthetically generated texts as training data. We automatically create positive and negative labels from the seed skill graph. The training data are encoded by pre-trained language models into embeddings and they are consumed by the downstream classification module to classify the relationships between skill pairs.