Personalized Product Description Generation With Gated Pointer-Generator Transformer

In the realm of e-commerce, where online shopping has become a staple of daily life, the generation of personalized product descriptions presents a unique challenge and opportunity for enhancing customer experience. Traditional retail interactions allow for personalized communication between salespersons and customers, ensuring that consumer needs are directly addressed. This level of personalization is harder to achieve online, where customers must navigate through generic, often lengthy product descriptions to make informed purchasing decisions. Recognizing the dual necessity of personalizing content to individual preferences while ensuring the descriptions remain faithful to the product's core attributes, this article introduces a novel approach, the gated pointer-generator transformer (GPGT). This framework is designed to bridge the gap between customer preferences and product features, enabling the generation of descriptions that are not only customized to the user's interests—such as emphasizing appearance for fashion-forward individuals or functionality for tech enthusiasts—but also accurately reflect the product's distinctive qualities, including brand names and technical specifications. GPGT leverages the select-attention mechanism combined with a Transformer encoder to capture the nuanced interactions between user attributes and product features, further refined by a copy mechanism during the decoding phase for the precise inclusion of specific product-related terms. Extensive experiments show that our framework substantially improves the quality of generation (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$+$</tex-math></inline-formula>10.6% on ROUGE-2 and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$+$</tex-math></inline-formula>15.9% on BLEU) while being more faithful to draw people's attention. The results on human evaluation, in terms of fluency, faithfulness, and personalization, also exhibit that descriptions generated by GPGT can be better accepted by real users.