[3J3-E-4] Robots and real worlds: Human Interactions

[3J3-E-4-01] Automatic Advertisement Copy Generation System from Images

〇Koichi Yamagata¹, Masato Konno¹, Maki Sakamoto¹ (1. The University of Electro-Communications)

Keywords:word2vec, automatic advertisement copy generation, RNN, automatic sentence generation

When we want to sell something, the presence of a good advertisement copy often affects sales. In this research, we develop an automatic advertisement copy generation system. Most existing systems only enter keywords, and the potential image of the product is not necessarily reflected by keywords only. We propose a method to generate advertisement copies using images as input to convey potential messages and the world view. This method uses Word2vec and color information of ad images, and both were confirmed to be effective by evaluation experiments. In the evaluation experiments, the mean score of the proposed method was significantly larger than 4 out of 7, and most of the subjects answered positively to ad copies of our method.