Method for Evaluating Quality of Automatically Generated Product Descriptions
Kenji Fukumoto, Risa Takeuchi, Akiyo Nadamoto · 2022
In recent years, deep learning accuracy has improved. Widely diverse content has been generated automatically. Along with studies of automatic content generation, difficulties associated with evaluation methods for the generated content have become important. The evaluation methods differ among the generated content types. As described herein, we propose evaluation methods aimed at motivating readers to purchase products through automatically generated product descriptions. Specifically, after extracting and modifying the 17 evaluation items from related research, we categorized evaluation items according to three axes signifying grammar, content, and attributes. Furthermore, to demonstrate the benefits of the proposed axes, we conduct user experiments of two types using the results obtained from deep learning of three types. Experiment 1 targets subjects who have not decided what to buy. Experiment 2 targets subjects with an image of a product that they want to buy. As results, we identified the benefits of the proposed method in both cases.