Unsupervised Product Title Optimization Based on Search Behavior Knowledge in E-commerce

Shu Liu,Zhiqiang Ye,Jian Liao,Jinxin Wu,Zhao Li

DOI: https://doi.org/10.1109/IJCNN54540.2023.10192008

2023-01-01

Abstract:Nowadays, E-commercial platforms such as Taobao and Amazon, provide massive products which greatly improve the convenience of online shopping. However, there have been many redundant and lengthy titles, which are misleading and hurt the experience of consumers. Under this background, the product title compression task has been proposed to generate readable short titles. The existing supervised models rely on paired corpora, which is often expensive to annotate. Whereas unsupervised methods are not adopted in the E-commercial scenario due to lack of domain knowledge. To overcome the above limitations, we propose an unsupervised title compression method based on search behavior graph. Specifically, we build the knowledge graph extracted from search queries to detect redundancy. Moreover, click behaviors are utilized to give a clue to recognize inaccurate information. Then we employ a novel graph-to-text model consisting of an edge-importance encoder and a length-countdown decoder. The encoder combines the title with both query knowledge and click behaviors, and the decoder generates the compressed title with the desired length. Experiments on datasets show that our model outperforms the state-of-the-art models on both automatic and human evaluations. Moreover, further analysis demonstrates that our method has the ability to detect redundancy and keep the appropriate information.