Automatic Text-based Clip Composition for Video News

News broadcasters must produce engaging video clips quicker than ever to ensure their successful positioning in the market. This is due, in part, to the growing number of news sources and changes in media consumption amongst target audiences. This evolution has amplified the need to quickly produce news clips, a requirement that remains at odds with the traditionally manual and time-consuming video editing processes. Besides advances in automating video news production, current systems are yet to meet the sufficient automation level and quality standards required for professional news broadcasting. Addressing this gap, we propose a novel transformer-based framework for automatically composing news clips to streamline the editing process. Our framework is predicated on a vision-language feature embedding mechanism and a cross-attention transformer architecture designed to generate multi-shot news clips semantically coherent with the editorial text and stylistically consistent with professional editing benchmarks. Our framework composes news clips with a length of 2 minutes from source material ranging from 20 minutes to 2 hours in less than 5 minutes using a single GPU. In our user study, target groups with different experience levels rated the generated videos on a 6-point Likert scale. Users rated the news clips generated by our framework with an average score of 4.13 and the manually edited news clips with an average score of 4.58.