Researchers from the University of Edinburgh have developed the GRAPHTRAILER model to improve movie trailer creation by identifying key storytelling points.

The model leverages screenplay information and a graph-based representation of movies, outperforming baselines and Transformer models through knowledge distillation.

The authors address model discontinuities using the Straight-Through Estimator and Gumbel-softmax reparametrization, with training involving binary cross-entropy loss for TP identification.

Self-supervised pre-training and sentiment flow analysis are crucial components, dividing trailers into sections of varying intensity to engage viewers effectively.

The study explores different training regimes and criteria for selecting shots based on narrative structure and sentiment, with combined information yielding the highest accuracy.

Despite efforts to avoid spoilers, the model prioritizes selecting exciting shots from the latest parts of the movie to create engaging trailers.

Future work will involve predicting fine-grained emotions in movies and developing new emotion datasets and detection models, showing promising results in trailer generation with potential for future improvements in spoiler identification techniques.