Vision Enhanced Generative Pre-trained Language Model for Multimodal Sentence Summarization

Multimodal sentence summarization (MMSS) is a new yet challenging task that aims to generate a concise summary of a long sentence and its corresponding image. Although existing methods have gained promising success in MMSS, they overlook the powerful generation ability of generative pre-trained language models (GPLMs), which have shown to be effective in many text generation tasks. To fill this research gap, we propose to using GPLMs to promote the performance of MMSS. Notably, adopting GPLMs to solve MMSS inevitably faces two challenges: 1) What fusion strategy should we use to inject visual information into GPLMs properly? 2) How to keep the GPLM′s generation ability intact to the utmost extent when the visual feature is injected into the GPLM. To address these two challenges, we propose a vision enhanced generative pre-trained language model for MMSS, dubbed as Vision-GPLM. In Vision-GPLM, we obtain features of visual and textual modalities with two separate encoders and utilize a text decoder to produce a summary. In particular, we utilize multi-head attention to fuse the features extracted from visual and textual modalities to inject the visual feature into the GPLM. Meanwhile, we train Vision-GPLM in two stages: the vision-oriented pre-training stage and fine-tuning stage. In the vision-oriented pre-training stage, we particularly train the visual encoder by the masked language model task while the other components are frozen, aiming to obtain homogeneous representations of text and image. In the fine-tuning stage, we train all the components of Vision-GPLM by the MMSS task. Extensive experiments on a public MMSS dataset verify the superiority of our model over existing baselines.