Leveraging language models for automated distribution of review notes in animated productions, Neurocomputing

Abstract

During the production of an animated film, professionals at the animation studio prepare thousands of notes. These notes describe improvements and corrections identified by supervisors and directors during daily meetings where the film’s progress is reviewed. After each meeting, these notes are manually distributed to the appropriate departments that need to address them. Due to the manual nature of this process, many notes are not assigned correctly, and the identified issues are not addressed, reducing the final quality of the film. This article describes and compares several approaches to automatically distribute notes using multi-label text classification with different language models (LM). Implemented methods include logistic regression models, encoder-only models such as the BERT family, and decoder-only models such as Llama 2 including fine-tuning and QLoRA techniques. Training and inference were conducted on a local RTX-3090. The results of the different techniques have been compared, achieving a maximum average accuracy of 0.83 and an f1-score of 0.89 with the fine-tuned Multilingual BERT model. This demonstrates the validity of these models for multi-label text classification, as well as their usefulness in a hitherto unexplored area such as animation studios.