LinkedIn's new AI suggests job search facets in real time
Over 80% of job queries are vague, but this AI disambiguates instantly.
Job seekers often start with vague queries — LinkedIn reports that over 80% of job-related searches use three or fewer keywords. To solve this, a team of researchers from LinkedIn has developed Dynamic Facet Suggestion (DFS), an interactive query refinement system that disambiguates user intent in real time. The system surfaces personalized semantic attributes (e.g., industry, seniority, skills) based on the joint user-query context. The core framework combines offline taxonomy curation, embedding-based retrieval of top-K candidates, and a distilled small language model (SLM) for candidate scoring. Real-time serving is optimized via pointwise single-token scoring with batching and prefix caching, ensuring low latency.
Offline evaluation showed high precision for generated suggestions, and online A/B tests confirmed significant improvements in both suggestion engagement and downstream job search outcomes. This approach effectively bridges the gap between short user queries and the rich, structured data in LinkedIn's job graph. The paper also offers a blueprint for any vertical search platform dealing with ambiguous, short queries — think e-commerce, travel, or real estate — where intent prediction is critical. By grounding suggestions in policy (business rules and constraints) rather than pure relevance, DFS balances personalization with fairness and scalability.
- Over 80% of LinkedIn job searches use ≤3 keywords, making intent inference difficult.
- DFS uses a retrieval-augmented framework with offline taxonomy, embedding retrieval, and distilled SLM scoring.
- Online A/B tests showed significant improvements in suggestion engagement and job search outcomes.
Why It Matters
A real-time AI system that turns vague job searches into precise, high-engagement queries.