Co-visitation Meets Token Alignment for Next Product Title Generation

Chenwang Wu, Leyan Deng, Zhihao Zhu, Defu Lian

Keywords: Session-based recommendation, title generation

Abstract: The title gives users a quick description of the product, which may help improve various downstream tasks of recommendations. Existing work utilizes language models with limited generation capabilities, which may lead to issues such as excessively brief and inefficient titles being produced. To this end, we proposed an efficient method to generate the next product title, consisting of the co-visitation recommendation module and the token alignment module. The co-visitation module gets the recommended title through a product co-visitation graph, while the token alignment module generates the final title by aligning the recommended title with the last interacted title to eliminate redundant tokens. Finally, we achieved high-quality titles and demonstrated competitive performance in the KDD CUP 2023 Tack3 competition (2nd place). Moreover, our method is very efficient and scalable, making it highly practical for large-scale systems.

Submission Number: 17