Deep Learning and Film History: Model Explanation Techniques in the Analysis of Temporality in Finnish Fiction Film Metadata
Chang Li-Hsin; Ginter Filip; Kanerva Jenna; Kiiskinen Harri; Salmi Hannu
We demonstrate the application of a deep-learning -based regressor, on a case study of predicting movie production year based on its plot summary. We show how the Integrated Gradients (IG) model explanation method can be used to attribute the predictions to individual input features and compare these to human-assigned attributions. Our purpose is to provide an insight into the application of modern NLP methods in the scope of a digital humanities research question, and test the model explanation techniques on a problem that is easy to understand, yet non-trivial for both humans and machine learning algorithms alike. We find that the model clearly outperforms non-expert human annotators, being able to date the movies well within the correct decade on average. We also demonstrate that the model-assigned attributions agree with those assigned by humans, especially for correct predictions.
- Rinnakkaistallenteet