Detecting emotional scenes using Semantic Analysis on Subtitles

In this project, our aim was to detect and clip emotional scenes from movies, using natural language processing techniques on their subtitles. Our approach involved the design of a naïve counting classifier, to which we gradually added semantic knowledge. With each increment, we analyzed the change in performance and made suitable modifications to our classifier. In our second approach, we trained a maximum entropy classifier on some handtagged subtitles and compared the performance of the MaxEnt classifier with different types of data and pre-learned emotional words. Using these classifiers, we then created an interactive tool for searching and watching 2-3 minute clips of movies. Subsequent sections of this paper describe our approaches in more detail.