How Alibaba’s Intelligent Writer Boosted Double‑11 Clicks with AI‑Generated Content

Introduction

Content‑driven experiences have become a key focus for Taobao in recent years. Various formats such as rich‑text product highlights in "Good Goods" and themed list pages ("Must‑Buy Lists") enrich the shopping experience by organizing products from multiple dimensions.

To scale this effort, Alibaba built the "Intelligent Writer" product, which combines massive data, human expertise, and knowledge inputs to automate and mass‑produce high‑quality content. The system now supports short copy, title generation, recommendation reasons, and visual list creation, with notable progress during the 2017 Double‑11 event.

Smart Benefit Points

During Double‑11, the Intelligent Writer generated billions of personalized short copy snippets for homepage and event entrances, dramatically increasing click‑through rates. Traditional benefit‑point creation was limited by manpower and often produced only one or two points per product, leading to mismatches and reduced diversity.

By integrating with recommendation and promotion teams, the system deployed intelligent benefit points across multiple entry points, as illustrated in the following images:

Online bucket tests showed double‑digit percentage lifts in click‑through rates compared to manually edited copy.

PairXNN Model Overview

The core of benefit‑point generation is the PairXNN model, which predicts the click probability of a product’s selling point for a given user. It consists of three main components:

User preference and product selling‑point semantic representations.

Multi‑level similarity modules (cosine similarity at the embedding level and bilinear similarity).

Additional engineered features (e.g., overlap between user preferences and selling points).

The model is trained and served on Alibaba’s internal XTensorflow platform.

Semantic Representation

Various architectures were evaluated for encoding user preference tags, including DNN, CNN, Gated CNN, self‑attention, and tailored attention. Gated CNN achieved the best trade‑off between performance and latency and was deployed for the Double‑11 rollout.

Multi‑Level Similarity Module

Two similarity calculations are performed:

Cosine similarity between user‑side and item‑side embedding vectors, followed by global max/min/average pooling.

Bilinear similarity using a trainable matrix M to map user vectors to the item space.

These similarity scores are concatenated and fed to subsequent layers, improving CTR over single‑level approaches.

Visual List Generation

Visual lists (themed product collections) are labor‑intensive to produce, especially during large promotions. The Intelligent Writer applied its NLG pipeline to generate short product recommendation reasons (40‑80 characters) and concise titles (<20 Chinese characters) for each list.

The generation pipeline uses an Encoder‑Decoder architecture with attention, coverage attention, context gates, beam search, and CNN encoders for title generation. Reinforcement learning (REINFORCE) aligns the training objective with the BLEU reward.

Sample generated titles and recommendation reasons demonstrate the system’s ability to produce diverse, appealing copy.

Title: "Sweat‑proof Hoodie for Youthful Energy"

Reason: "The V‑neck design reveals a sexy collarbone, while the color‑blocked stitching adds a fashionable flair."

Future Work

While the Intelligent Writer achieved promising results, several challenges remain:

Better evaluation metrics beyond BLEU and manual assessment.

Incorporating richer inputs such as product images and user reviews.

Improving model interpretability and addressing bad‑case analysis.

Enhancing diversity and accuracy of generated text, and exploring transferability to low‑resource domains.

References

[1] Severyn & Moschitti, "Learning to rank short text pairs with convolutional deep neural networks," SIGIR 2015. [2] Vaswani et al., "Attention is all you need," arXiv 2017. [3] Dauphin et al., "Language modeling with gated convolutional networks," arXiv 2016. [4] Luo et al., "Understanding the effective receptive field in deep convolutional neural networks," NIPS 2016. [5] Bahdanau et al., "Neural Machine Translation by Jointly Learning to Align and Translate," 2015. [6] Rush et al., "A neural attention model for abstractive sentence summarization," arXiv 2015. [7] Kim, "Convolutional neural networks for sentence classification," arXiv 2014. [8] Tu et al., "Modeling coverage for neural machine translation," arXiv 2016. [9] Tu et al., "Context gates for neural machine translation," arXiv 2016. [10] Ranzato et al., "Sequence Level Training with Recurrent Neural Networks," ICLR 2016. [11] Papineni et al., "BLEU: a method for automatic evaluation of machine translation."