Evaluating Face Tracking for Political Analysis in Japanese News Over a Long Period of Time

TV news is the major source of political information for most of the people, exercising a strong influence on public opinion. Japanese news media try to carefully balance politicians’ representation, but it is important to empirically examine this balance longitudinally. However, it is a tedious, and in many cases impossible task to achieve manually, especially when the news archive covers over a decade. We therefore rely on automatic procedures to do it computationally rather than manually. In this paper, we compare the two face tracking methods as well as a traditional text-based method by using the same dataset of Japanese broadcasting news spanning over a decade. We evaluate the three methods against a manually curated random sample of NHK’s News 7, the flagship news program of the Japanese public broadcasting. The first tracking method is inherited from previous works used on the same dataset and based on traditional Viola-Jones detections and VGGFace for embeddings. Our second method uses modern deep learning techniques with MTCNN for face detection, and ResNet50 trained with VGGFace2 for embeddings. We not only demonstrate that our modern implementation outperforms the two other methods, but also discuss implications and application for social scientific studies.