Abstract
In product description generation (PDG), the user-cared aspect is critical for the recommendation system, which can not only improve user's experiences but also obtain more clicks. High-quality customer reviews can be considered as an ideal source to mine user-cared aspects. However, in reality, a large number of new products (known as long-tailed commodities) cannot gather sufficient amount of customer reviews, which brings a big challenge in the product description generation task. Existing works tend to generate the product description solely based on item information, i.e., product attributes or title words, which leads to tedious contents and cannot attract customers effectively. To tackle this problem, we propose an adaptive posterior network based on Transformer architecture that can utilize user-cared information from customer reviews. Specifically, we first extend the self-attentive Transformer encoder to encode product titles and attributes. Then, we apply an adaptive posterior distillation module to utilize useful review information, which integrates user-cared aspects to the generation process. Finally, we apply a Transformer-based decoding phase with copy mechanism to automatically generate the product description. Besides, we also collect a large-scare Chinese product description dataset to support our work and further research in this field. Experimental results show that our model is superior to traditional generative models in both automatic indicators and human evaluation.
Introduction
In E-commerce, a compelling product description does more than list technical specifications; it highlights "user-cared aspects"—the specific details like "breathable fabric" or "easy to carry" that resonate with potential buyers. Automated systems typically generate descriptions from dry product titles and attribute tables, often missing the persuasive "soul" found in customer feedback. While customer reviews are a rich source of these insights, they present a "cold start" problem for new or "long-tailed" products that have yet to receive feedback. This research addresses how to inject review-inspired nuances into automated descriptions without requiring reviews to be present at the time of generation.
The Adaptive Posterior Distillation Transformer (APDT)
To bridge the gap between information-rich reviews and review-less new products, we propose the Adaptive Posterior Distillation Transformer (APDT). The model utilizes a teacher-student framework to transfer knowledge from a review-aware "teacher" to a review-less "student."
Methodology
The APDT architecture is built on three main pillars:
- Encoder Layer: A Transformer-based encoder that separately processes the product title, its attributes, and (during training only) customer reviews. The authors found that distilling from deeper layers (5th or 6th) of the teacher's encoder was most effective for capturing semantic nuance.
- Adaptive Posterior Distillation: This is the core innovation. During training, a "Teacher" model has access to the actual customer reviews (posterior information) and learns a representation highlighting user-cared aspects. A "Student" model, which only sees the title and attributes, is trained to minimize the KL Divergence between its representation and the teacher's. This process distills the knowledge of what users care about into the student's generation logic.
- Decoder with Copy Mechanism: To maintain factual accuracy, the model employs a pointer-generator style copy mechanism. This allows it to directly copy technical details from the input while generating persuasive, review-inspired prose.
- Coherence-Enhanced Loss ($L_{CoE}$): A custom loss function rewards the generation of informative, less-frequent words that are highly coherent with the ground-truth description.
Dataset and Task Illustration
We introduced a large-scale benchmark dataset, JDPDG, collected from JD.com. It contains 345,799 pairs of item content and descriptions across categories such as Shoes & Clothes, Digital, and Homing. Each entry includes a title, key-value attributes, and high-quality customer reviews.
Experimental Results
The APDT model was evaluated against several state-of-the-art baselines, including KOBE and standard Transformer models.
- Automatic Metrics: In the "Shoes & Clothes" category, APDT achieved a ROUGE-L of 20.41 and a BLEU score of 10.36, significantly outperforming the previous best model, KOBE (19.07 and 9.27, respectively).
- Human Evaluation: In blind tests, human evaluators rated APDT higher in both Correctness (2.91/4.0) and Coherence (3.02/4.0) compared to all baselines.
- Qualitative Analysis: Case studies demonstrate that APDT successfully moves beyond simple attribute listing. For instance, while baseline models might only mention a product's color, APDT adds contextually relevant phrases like "very convenient for office" or "runs very smoothly," learned from its review-based training.
Conclusion
APDT effectively addresses the cold-start problem in product description generation by leveraging a posterior distillation framework. By treating reviews as a "soft teacher" during training, the model learns to anticipate and highlight features that customers value most. Future research will explore incorporating Question-Answering (QA) data and multi-task learning to further refine the generation of user-centric product content.
Summary updated