Generating enduring advertisements that attract shoppers is a nontrivial challenge for any e-commerce website, particularly maintaining creative quality at scale. To address this, we propose a programmatic solution to generate product advertising headlines using retail content. Our approach leverages state-of-the-art reinforcement learning (RL) policy gradient methods on transformer-based masked language models to create advertising headlines. Specifically, we generate headlines by jointly conditioning on multiple products that a seller wishes to advertise.
Our experiments demonstrate that our method outperforms existing transformer and LSTM + RL methods in overlap metrics and quality audits. Furthermore, the headlines generated by our model surpass human-submitted headlines in terms of both grammar and creative quality as determined by audits.
Blogger's Review: This paper presents an innovative approach to applying reinforcement learning in ad headline generation, showcasing the potential of machine-generated content in creative fields. As technology advances, the automation of advertising creation is expected to increase, especially with the ability to generate context-sensitive headlines across multiple products, enhancing both relevance and appeal.