Automatic Product Copywriting for E-Commerce: The APCG System and Its AI Innovations

Recently, the 36th AAAI International Conference announced that JD.com’s Retail Search and Recommendation Platform’s "Automatic Product Copywriting for E‑Commerce" (APCG) won the AAAI 2022 Innovation Application Award, one of eight global AI innovation awards.

The APCG system provides automated copy generation for multiple JD platforms, including the Discover Good Products channel, JD Books, JD Search, Global Shopping, Home & Living, and the live‑streaming platform.

APCG consists of two main parts: (1) a natural language generation model built from a Transformer‑Pointer Network combined with a pretrained sequence‑to‑sequence model trained on JD’s internal data, and (2) a copy quality control module that includes automatic detection, filtering, and manual review.

Because product categories (e.g., clothing, phones, sportswear) have vastly different features and data attributes, training a separate model for each category would be costly. The system therefore uses a shared encoder‑decoder Transformer backbone (Vaswani et al., 2017) with a copy mechanism in the encoder to capture product information, while a pretrained copywriting generation model (PCG) handles low‑resource or newly released items.

Two novel pretraining objectives were designed: (1) a sentence reordering task that shuffles sentences and forces the model to reconstruct the original order, and (2) a pseudo‑summary generation task that selects a quarter of the clauses as targets and uses the remaining three‑quarters as input, encouraging the model to learn both extraction and generation.

After pretraining, the model is fine‑tuned on a dataset covering all product categories and deployed across JD’s various platforms, including a real‑time writing assistant for merchants on the live‑streaming channel.

By September 2021, APCG had generated 2.53 million product copies, increasing click‑through rate (CTR) by 4.22%, conversion rate (CVR) by 3.61%, and gross merchandise volume (GMV) by 213.42%. The system can produce up to 209,000 new materials per week, equivalent to the output of 1,000 human creators, delivering substantial labor cost savings and demonstrating the practical impact of AI‑driven natural language generation in e‑commerce recommendation platforms.