Intelligent Search Strategy for 58 Recruitment: Breaking Category Constraints and Building a Smart Recall Framework

The previous 58 recruitment search relied on a strict category hierarchy, causing many relevant job posts to be missed and leading to low commercial fill rates. To address this, the team reconstructed the retrieval logic, removing category constraints and exposing all relevant posts under the primary employment category.

The new intelligent search strategy consists of five modules: Query Understanding, Document (Doc) Understanding, Recall, Coarse Ranking, and Fine Ranking. Query rewriting expands and normalizes user queries, while intent recognition maps queries to structured tags such as position, work nature, and benefits.

Doc understanding filters out misleading or irrelevant text in job posts, using semantic models (BERT) to identify the true hiring position and discard noisy content.

Recall is performed through three parallel channels: a Text channel (query rewriting + doc understanding), an Intent channel (structured tag matching), and a Vector channel (dual‑tower embedding similarity). The vector channel captures semantically similar posts that lack explicit keyword overlap.

Model training leverages both behavior‑derived positive samples (shared clicked posts) and manually annotated data, with techniques such as random job substitution, key‑information swapping, and hard‑sample weighting to improve robustness. The dual‑tower architecture balances retrieval efficiency with semantic matching quality.

Online results show a 17% accuracy boost in the first phase and 10‑20% improvements across downstream products, with vector recall contributing 15‑20% growth in PVR and ASN metrics while maintaining overall precision.

Future work will focus on deeper intent mining for ambiguous queries (e.g., "summer job"), personalized query rewriting, and enhancing long‑text modeling with attention mechanisms.

Author: Cao Ranran, Senior Algorithm Engineer at 58.com, MSc from the University of Manchester, responsible for commercial search algorithm development.