Novel ad text generation for real estate

Abstract

Transformer architectures have brought unreasonable progress to natural language generation in recent years. However, generating high quality and relevant advertising text is difficult. Inspired by Laban et. al., reinforcement learning is used in this paper to encourage a generative pretrained transformer to generate text that adheres to the norms and criteria of real estate advertising on social media. Using a proprietary dataset of real estate listings and ad copies, it is found that the architecture proposed by Laban et. al. is sufficiently modular and customizable for use in other domains, albeit with a modification to the pretraining procedure. Evidence is presented that the model’s ability to embed desired information into the summary is contingent on the information being uncommon or unique.