Ad Headline Generation using Self-Critical Masked Language Model

The authors propose a programmatic solution for generating product advertising headlines by applying Reinforcement Learning policy gradient methods to Transformer-based Masked Language Models. This approach jointly conditions on multiple products that a seller wishes to advertise.

• The method utilizes RL Policy gradient methods on Transformer based Masked Language Models.
• It creates advertising headlines by jointly conditioning on multiple products.
• The model outperforms existing Transformer and LSTM + RL methods in overlap metrics and quality audits.
• Model-generated headlines surpass human submitted headlines in grammar and creative quality according to audits.

The authors consider this significant because it provides a scalable way to pass the creative quality bar for e-commerce websites, which is difficult to achieve manually at large scale.