Abstract
A compelling product description on e-commerce platforms (e.g., Amazon) is vital in explaining the untouchable product and increasing consumers’ purchase rate. However, hand-written product descriptions for each category of products is highly time-consuming and not professional due to the lack of related marketing knowledge. Prior works adopt either predefined templates or data-driven models to automatically generate the personalized product description, but the quality (e.g., readability, flexibility) of generated text is always constrained by the lack of personalized production description training samples. To further improve the product description quality, we propose a personalized product description generation model named CrowdDepict focusing on what proper permutation of attribute words should be taken to generate the description and how to describe the attribute words. Particularly, CrowdDepict integrates **an Attribute Permutation-insensitive Encoder **to enable the model to generate logical description with an appropriate attribute keywords organization without requiring a re-organized input attribute keywords and a Crowd Intelligence-aware Comment Encoder to capture crowd intelligence about how the attributes of products are described in real-world user comments. Experiment results demonstrate that CrowdDepict outperforms the baseline on various metrics, especially an improvement of 34% over state-of-the-art relative to BLEU, which shows that our model can generate personalized product description that consists of correct product attributes of consumer interests and the necessary product information.