Deep Semantic Model Exploration and Application in 58 Search

The talk begins with an introduction to 58 Search, describing its offline index construction and online retrieval pipeline, which consists of multiple layers (L1–L4) handling recall and ranking.

Traditional term‑match techniques work well for literal query‑document matching but struggle with semantic similarity when no exact term overlap exists; the presentation motivates the need for deep semantic models.

Three categories of deep semantic models are explored: fully interactive models (e.g., BERT with query‑document concatenation), dual‑tower (representation) models that compute document embeddings offline, and semi‑interactive (weakly interactive) models such as Poly‑Encoder that balance accuracy and latency.

Data preparation involves sampling positive and negative query‑document pairs from logs and manual annotation, with various sampling rules to improve label quality.

Experiments with BERT‑based interactive models show high accuracy but high inference cost, suitable only for the final L4 ranking layer. Dual‑tower models offer fast inference and can be applied to L2–L4, though with lower matching quality; extensive tuning (different loss functions, knowledge distillation from interactive models) narrows the performance gap.

Semi‑interactive models, especially Poly‑Encoder, achieve performance close to interactive models while maintaining acceptable latency, making them viable for L2 and L3 stages.

Application results demonstrate that adding semantic features to L3 and L4 ranking (e.g., XGBoost, DIEN) improves AUC, CTR, and CVR. In recall, semantic similarity is incorporated at the L2 layer to supplement strict term‑match, reducing no‑result cases and increasing relevant hits.

Future work includes further online deployment of semi‑interactive models, exploring higher‑accuracy interactive models for ranking, expanding semantic recall to more stages, and optimizing vector dimensions and serving infrastructure.