Distinctive Slogan Generation with Reconstruction
E-commerce sites include advertising slogans along with information regarding an item. Slogans can attract viewers’ attention to increase sales or visits by emphasizing advantages of an item. The aim of this study is to generate a slogan from a description of an item. To generate a slogan, we apply an encoder–decoder model which has shown effectiveness in many kinds of natural language generation tasks, such as abstractive summarization. However, slogan generation task has three characteristics that distinguish it from other natural language generation tasks: distinctiveness, topic emphasis, and style difference. To handle these three characteristics, we propose a compressed representation–based reconstruction model with refer–attention and conversion layers. The results of the experiments indicate that, based on automatic and human evaluation, our method achieves higher performance than conventional methods.
PDF Abstract
