Generating keywords-aware advertising texts assets using fine-tuned GPT-2

Abstract

It is necessary to reduce the workload of ad creators because it costs a great deal of time and money to manually create appropriate and appealing advertising messages for an ever-increasing number of products and services. In addition, since presenting the same advertisement again significantly decreases user response, it is necessary to create a variety of advertisements. Although research is conducted to reduce the burden on ad creators and avoid ad exhaustion by automatically generating advertisement using AI, the generation of variational advertisements is a challenge. We propose a method that focuses on advertising keywords using pre-trained on large-scale Japanese data transformer in order to generate various advertising messages that take into account the prerequisite knowledge about the product. We aim to generate various types of advertisement text considering the characteristics and prerequisite knowledge of the advertisement by extracting the keywords in the advertisement text and learning by conditioning on the keywords and the advertisement text. For comparison, we performed automatic evaluation of BLEU, distinct, and keyword occurrence rate, and manual evaluation of fluency and relevance. The results show that the proposed method generates various types of advertisements that capture the characteristics of keywords even with a small amount of data.