Language Models for Automatic Distribution of Review Notes in Movie Production

During the several years of production of an animated movie, review meetings take place daily, where supervisors and directors generate text notes about fixes needed for the movie. These notes are manually assigned to artistic departments for them to fixed. Being manual, many notes are not properly assigned and are never fixed, lowering the quality of the final movie. This paper presents a proposal for automating the distribution of these notes using multi-label text classification techniques. The comparison of the results obtained by fine-tuning several transformer-based language models is presented. A highest mean accuracy of 0.776 is achieved assigning several departments to each of the review notes in the test set with a BERT Multilingual model. A mean accuracy of 0.762 was reached in just 10 epochs and 10 minutes of training on an RTX-3090 with a DistilBERT transformer model.

GARCÉS Diego;  SANTOS Matilde;  FERNANDEZ LLORCA David; 

2023-12-07

SPRINGER VERLAG

JRC134507

1611-3349 (online),   

https://link.springer.com/chapter/10.1007/978-3-031-48232-8_23,    https://publications.jrc.ec.europa.eu/repository/handle/JRC134507,   

10.1007/978-3-031-48232-8_23 (online),