Multi-modal product title compression

Product title generation in e-commerce is a challenging task, which involves modeling multi-modal resources, i.e., textual descriptions and visual pictures, and comprising a sequence of words with proper ordering. Although myriad researches have studied this task and prompting progress has been made, there still exists a noticeable gap between generated titles and the requirements on mobile devices, especially considering the limited screen size. Towards filling this gap, we collect a large dataset from real e-commerce platforms to investigate compressing product titles for mobile devices, namely product title compression. We also propose a novel title compression model which takes the advantages of reinforcement learning and multi-modal resources. In doing so, our model is capable of retaining vital information in titles and improving the readability of generated titles. Experimental results demonstrate that our proposed method outperforms the state-of-the-art methods by a large margin on the automatic evaluation.