LinkedIn Unifies Job Understanding with Small Language Models

LinkedIn has introduced a new unified semantic modeling framework designed to improve its large-scale job understanding capabilities. This system is crucial for connecting talent with opportunities by converting noisy, unstructured job postings into standardized and derived job attributes that power numerous LinkedIn features. The core of the framework is a fine-tuned open-source small language model (SLM). This SLM is trained on synthetic tasks augmented with reasoning traces, enabling robust zero-shot generalization for both taxonomy-guided classification and taxonomy-agnostic entity extraction in job understanding contexts. Further enhancing the system, a multi-adapter architecture with attribute grouping facilitates efficient, task-specific adaptation and streamlines model management across diverse downstream attributes. Offline evaluations and online A/B tests have demonstrated significant performance improvements and reduced operational complexity, offering practical insights for building industrial-scale text understanding systems.