Intelligent Recommendation Selling Point Generation: Architecture, Core AI Techniques, Model Development, and Product Impact

Introduction – JD's intelligent merchant recommendation selling points are NLP‑driven product descriptions that enhance e‑commerce recommendation explanations and drive user engagement.

Technical Background – Traditional product copy is manually written; modern approaches use text summarization and natural language generation, including template‑based, deep neural network‑based, knowledge‑enhanced, and pattern‑controlled methods.

System Architecture – The workflow starts with the SOA/Mixer module, passes user and product data to Broadway (frontend), then to Index, followed by AI‑flow for recall and ranking, and finally to the selling‑point module which extracts and personalizes highlights for display.

Core AI Techniques – Two‑stage processing: (1) coarse filtering using a self‑adversarial BERT classifier to select high‑quality candidate sentences; (2) generation using a Transformer encoder‑decoder combined with a Pointer‑Generator to produce concise selling points; (3) fine‑grained filtering with a recursive sharpening BERT model to rank final outputs.

Personalized Distribution – User interest vectors are built via word2vec‑style embeddings of product terms weighted by user preferences; selling‑point vectors are similarly embedded, and similarity (e.g., cosine) determines the most relevant highlight for each user.

Model Development & Practice – Experiments compare material sources (user reviews vs. OCR‑extracted product details), showing both improve click‑through and UV metrics. Online monitoring uses business KPIs (exposure, click value, dwell time) to filter low‑quality points, while offline fine‑tuning with high‑quality samples further refines the BERT scorer.

Product Deployment & Results – The system now generates selling points for billions of SKUs across 62 categories, improving click‑through rates by ~2% and dwell time by ~0.3%, and is integrated into multiple JD platforms (main site, JD Joy, fast‑track, etc.).

Conclusion – Automated, AI‑driven selling‑point generation provides scalable, personalized product explanations that boost e‑commerce performance.