A novel multi-modal feature extraction system for news video

Abstract

Advancements in artificial intelligence technology are leading to significant transformations in the generation of news. This article presents the News Video Style Feature Automatic Extraction System (NVFAE), which leverages state-ofthe-art open-source technologies such as Automatic Speech Recognition, Optical Character Recognition, and Natural Language Processing. The NVFAE extracts news video style features from multiple modalities: textual features, visual features, and delivery features. The system uses automated annotation to extract five distinct style features from news videos: high-frequency words, overall tense, video color tone, main character, and video length. 38 BBC videos about the topic of "Turkey earthquake" from the YouTube platform are chosen for testing NVFAE, whose five features can be all successfully extracted. Experimental results show that the system proposed in this paper can provide a digital and structured way to explore the narrative characteristics of news videos, only with a mouse click and style features of news videos can be obtained in batches. NVFAE is aimed to overcome the difficulties of manual annotation for video news, as well as supply news workers with valuable references to study the news styles and content planning.