Artificial intelligence (AI) machine learning is being applied across nearly every industry, including motion-picture entertainment. This week researchers based in Southern California announced a new AI machine learning tool that can rate movie content in seconds in advance of any film production.
Victor Martinez at the University of Southern California (USC) was the lead researcher in the AI study that debuts this month at The Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. The other USC researchers include Krishna Somandepalli and Shrikanth Narayanan, working in collaboration with Yalda Uhls at the University of California, Los Angeles (UCLA). Together the team created a deep learning model to analyze and predict movie ratings based on the language data in the movie script.
“Computational methods that identify portrayals of risk behaviors from audio-visual cues are limited in their applicability to films in post-production, where modifications might be prohibitively expensive,” the researchers wrote. “To address this limitation, we propose a model that estimates content ratings based on the language use in movie scripts, making our solution available at the earlier stages of creative production.”
The process of rating movies is largely a manual, time-consuming process performed by human reviewers watching films. Movie ratings impact audience reach, and therefore have a significant fiscal impact on its overall potential revenue. Films are classified based on content that may be deemed unsuitable for children or adolescents such as substance abuse, violence, sex, and profanity.
“The present work, to the best of our knowledge, is the first to model the co-occurrence of risk behaviors from linguistic cues found in movie scripts,” reported the researchers. “Our proposed model is a multi-task approach that predicts a movie script’s violent, sexual, and substance-abusive content from vectorial representations of the character’s utterances.”
The researchers created a machine learning model that learns how to map sequences of character utterance representations to overall movie ratings, rather than a word-based representation. An utterance can be one or more words, or a sentence that is preceded and followed by silence. The representation comprises of both sentiment and semantics. The model inputs a sequence of character utterance representations and produces content rating predictions. The model was trained using over 980 movie scripts.
The researchers’ model, MovieBERT, is a bidirectional encoder representation from transformers (BERT), with 12 transformer layers that "learn a 768-dimensional representation of a movie script.” MovieBERT is optimized to perform next-sentence prediction and masked language modeling.
The method uses a recurrent neural network (RNN), a class of AI neural networks commonly used in natural language processing and speech recognition that allow previous outputs as inputs while having hidden states. For example, Amazon Alexa and Google Voice Search use recurrent neural networks.
"Our model significantly improves the state-of-the-art by adapting novel techniques to learn better movie representations from the semantic and sentiment aspects of a character’s language use, and by leveraging the co-occurrence of risk behaviors, following a multitask approach,” reported the researchers. "MovieBERT achieves 96.5% accuracy on the next sentence prediction task, and a 65.9% accuracy on the masked language model—an absolute improvement from the BERT-base model of 24.5% and 12.43%, respectively."
By applying the pattern-recognition and predictive capabilities of AI deep learning, the motion picture industry now has a new way to evaluate the potential audience impact of a movie before production, in order to save time and costs in the future.
Copyright © 2020 Cami Rosso All rights reserved.