MultiSentimentArcs: a novel method to measure coherence in multimodal sentiment analysis for long-form narratives in film

Affective artificial intelligence and multimodal sentiment analysis play critical roles in designing safe and effective human-computer interactions and are in diverse applications ranging from social chatbots to eldercare robots. However emotionally intelligent artificial intelligence can also manipulate, persuade, and otherwise compromise human autonomy. We face a constant stream of ever more capable models that can better understand nuanced, complex, and interrelated sentiments across different modalities including text, vision, and speech. This paper introduces MultiSentimentArcs, combination of an open and extensible multimodal sentiment analysis framework, a challenging movie dataset, and a novel benchmark. This enables the quantitative and qualitative identification, comparison, and prioritization of conflicting sentiments commonly arising from different models and modalities. Diachronic multimodal sentiment analysis is especially challenging in film narratives where actors, directors, cinematographers and editors use dialog, characters, and other elements in contradiction with each other to accentuate dramatic tension. MultiSentimentArcs uses local open-source software models to democratize artificial intelligence. We demonstrate how a simple 2-step pipeline of specialized open-source software with a large multimodal model followed by a large language model can approximate video sentiment analysis of a commercial state-of-the-art Claude 3 Opus. To the best of our knowledge, MultiSentimentArcs is the first fully open-source diachronic multimodal sentiment analysis framework, dataset, and benchmark to enable automatic or human-in-the-loop exploration, analysis, and critique of multimodal sentiment analysis on long-form narratives. We demonstrate two novel coherence metrics and a methodology to identify, quantify, and explain real-world sentiment models and modalities. MultiSentimentArcs integrates artificial intelligence with traditional narrative studies and related fields like film, linguistic and cultural studies. It also contributes to eXplainable artificial intelligence and artificial intelligence safety by enhancing artificial intelligence transparency in surfacing emotional persuasion, manipulation, and deception techniques. Finally, it can filter noisy emotional input and prioritize information rich channels to build more performant real-world human computer interface applications in fields like e-learning and medicine. This research contributes to the field of Digital Humanities by giving non-artificial intelligence experts access to directly engage in analysis and critique of research around affective artificial intelligence and human-AI alignment. Code and non-copyrighted data will be available at https://github.com/jon-chun/multisentimentarcs .