Interactive BERT for Relevance in Health E‑commerce Search

Health Search Business and Technical Overview

The health e‑commerce search primarily operates on Taobao, covering drug, medical device, and health supplement listings. Relevance is critical due to professional terminology in user queries, requiring precise matching between intent and product attributes.

Interactive BERT Algorithm Exploration

Three main challenges were identified: (1) high relevance requirements in a vertical domain, (2) scarcity of high‑quality labeled samples, and (3) the need for low‑latency online inference. To address these, three optimization directions were proposed: enhancing key‑attribute understanding, constructing high‑quality samples, and strengthening dual‑tower semantic representation.

Model Design

Text feature extraction for queries and products, including fine‑grained entity parsing (e.g., generic name, brand, dosage).

Domain‑specific sample generation using click logs, exposure‑bias correction, and entity/relationship knowledge.

Interactive BERT architecture with keyword embedding that flags tokens representing key attributes, an additional feature‑matching network on top of the 11th transformer layer, and concatenation with the 12th layer before the final fully‑connected layer.

The enhanced model achieved an AUC of 0.919 on a manual test set, and after further pre‑training with a click‑prediction task, the AUC rose to 0.927.

Model Application Practice

Due to the large parameter count, the interactive model was distilled into a dual‑tower model using 2 billion domain samples. The distilled model retains most of the accuracy (AUC loss of only 0.6 points) while reducing inference latency to under 10 ms.

Online A/B testing showed a 2.6‑point increase in manual relevance rating, and over 3 % uplift in order volume and revenue for both Alibaba Health Pharmacy and Tmall Good Medicine platforms.

Q&A Highlights

Long‑tail and zero‑click items are covered by leveraging a year‑long search log.

Relevance standards combine professional medical criteria with user behavior data.

Negative samples are derived from exposure‑without‑click logs after filtering by text and entity scores.

Unigram and bigram embeddings are obtained via hash embeddings, avoiding a predefined vocabulary.

Knowledge distillation uses a large teacher model to generate soft labels for the smaller dual‑tower student.

Overall, the interactive BERT approach significantly improves relevance in health e‑commerce search while maintaining production‑grade efficiency.