Optimizing E-commerce Product Copy Generation: Challenges, Framework, and System Practices

The talk, presented by JD algorithm engineer Chen Hongshen, outlines the importance of product copy in e‑commerce and the two main technical challenges of encoder‑decoder based text generation: unreliable output quality and data collapse leading to generic, low‑diversity copy.

To address these issues, a three‑component solution is proposed: a copy‑cleaning system that filters and extracts high‑quality textual fragments from user reviews and product details, a copy generation system based on transformer‑pointer models enhanced with large‑scale pretrained language models, and a copy quality evaluation system that filters out low‑quality results before online deployment.

The framework is illustrated with diagrams and applied to both short and long copy use cases. Case studies demonstrate how the cleaning system uses adversarial fine‑tuning and cascade filtering, how the generation system incorporates massive pretrained models, posterior distillation for long‑tail items, reference‑template copying, and data augmentation via synonym replacement and back‑translation.

Further optimizations include multi‑model cascades for higher precision, knowledge distillation from popular to long‑tail products, and weighting strategies for augmented training data to mitigate noise.

Finally, the article references two published papers (AAAI21 and SIGIR20) that detail the underlying research, and concludes with acknowledgments and promotional information for related data and algorithm resources.