NLP Algorithm Practices in Alibaba's Brand Advertising

The talk, delivered by senior Alibaba algorithm expert Xiao Guorui, introduces the overall landscape of Alibaba Mama's brand advertising, covering product matrices, resource channels, and the deterministic business model that emphasizes pre‑order inventory locking, contract‑based budgeting, and guaranteed delivery metrics such as UV, PV, and reach.

The technical architecture is divided into four stages—query estimation, order placement, real‑time delivery, and post‑delivery evaluation—highlighting the need for precise targeting, inventory management, flow allocation, and automated reporting, with distinct pipelines for brand display and brand search ads.

Two core NLP tasks are addressed: (1) brand intent recognition, which classifies whether a user query expresses a desire to purchase a specific brand’s product, and (2) short‑text relevance, which scores the semantic match between a query and an advertised item using a learning‑to‑rank framework.

For brand intent recognition, the solution follows a two‑step pipeline: blocking to generate candidate pairs and matching to determine intent. Model evolution progressed from an Entailment‑based BERT fine‑tune to a Multi‑Choice BERT approach, incorporating segmented features (query text, clicked item titles, category predictions) and handling multiple candidate brands.

Short‑text relevance is modeled as a ranking problem with relevance levels 0‑5. Initial DSSM models showed promising metrics but suffered from Gaussian‑like score distributions that destabilized threshold decisions. Upgrading to a pointwise binary classifier with a Compare‑Aggregate architecture produced a bimodal (Bernoulli) distribution, reducing threshold sensitivity and improving AUC/NDCG.

Evaluation results demonstrate that the Multi‑Choice and Compare‑Aggregate models consistently outperform earlier baselines, while insights reveal the importance of fitting Bernoulli distributions for deterministic ad delivery and the superiority of interaction‑based models over representation‑only approaches.

The summary concludes with practical lessons: leveraging weak supervision and data augmentation for large‑scale labeling, emphasizing problem definition as a core skill for applied algorithm engineers, and advocating incremental, low‑cost iterative development for industrial AI deployments.