Towards story-based classification of movie scenes

• Published in: PloS one 2020-02, Vol.15 (2), p.e0228579-e0228579
• Main Authors: Liu, Chang, Shmilovici, Armin, Last, Mark
• Format: Article
• Language: English
• Subjects: Algorithms; Analysis; Annotations; Biology and Life Sciences; Cartoons; Classification; Computer and Information Sciences; Computers; Datasets; Divergence; Emotions; Engineering and Technology; Ground truth; Guidelines; Humans; Labels; Learning algorithms; Machine Learning; Medicine and Health Sciences; Models, Psychological; Motion pictures; Motion Pictures - classification; Music; Narratives; Physical Sciences; Probability distribution; Recommender systems; Research and Analysis Methods; Screenplays; Similarity; Social Sciences; Structure (Literature); Television; Television programs; Video data
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0228579

Humans are entertained and emotionally captivated by a good story. Artworks, such as operas, theatre plays, movies, TV series, cartoons, etc., contain implicit stories, which are conveyed visually (e.g., through scenes) and audially (e.g., via music and speech). Story theorists have explored the structure of various artworks and identified forms and paradigms that are common to most well-written stories. Further, typical story structures have been formalized in different ways and used by professional screenwriters as guidelines. Currently, computers cannot yet identify such a latent narrative structure of a movie story. Therefore, in this work, we raise the novel challenge of understanding and formulating the movie story structure and introduce the first ever story-based labeled dataset-the Flintstones Scene Dataset (FSD). The dataset consists of 1, 569 scenes taken from a manual annotation of 60 episodes of a famous cartoon series, The Flintstones, by 105 distinct annotators. The various labels assigned to each scene by different annotators are summarized by a probability vector over 10 possible story elements representing the function of each scene in the advancement of the story, such as the Climax of Act One or the Midpoint. These elements are learned from guidelines for professional script-writing. The annotated dataset is used to investigate the effectiveness of various story-related features and multi-label classification algorithms for the task of predicting the probability distribution of scene labels. We use cosine similarity and KL divergence to measure the quality of predicted distributions. The best approaches demonstrated 0.81 average similarity and 0.67 KL divergence between the predicted label vectors and the ground truth vectors based on the manual annotations. These results demonstrate the ability of machine learning approaches to detect the narrative structure in movies, which could lead to the development of story-related video analytics tools, such as automatic video summarization and recommendation systems.