Length Controllable Model for Images and Texts Summarization

With the proliferating of multimedia data, controllable summarization generation has become a hot spot in Artificial Intelligence Content Generation area. However, some traditional summarization approaches are unable to accurately control the length of summarization resulting in content that is either overly verbose or lacking in information, which makes it difficult to satisfy users’ diverse reading demands thereby. In this paper, we propose length controllable approach for images and texts summarization. Our approach 1). puts preset character length information into a loop decoding unit as input, 2). combines with attention mechanism to control the length of the summarization, while using visual vectors to select the image with the highest correlation. In the task of length contrallable images and texts summarization, our approach outperforms models on MSMO dataset by the implementation of length control settings.