Spoiler Detection with Deep Learning in Turkish-Written Movie Reviews
33rd Conference on Signal Processing and Communications Applications-SIU-Annual, İstanbul, Turkey, 25 - 28 June 2025, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/siu66497.2025.11112173
- City: İstanbul
- Country: Turkey
- Istanbul University Affiliated: No
Abstract
Before deciding to watch a movie, it is usually consulted the reviews of those who have already watched that movie. However, reviews containing spoilers are a serious problem for those who have not yet watched the movie. This study is the first research on automatic spoiler detection on movie reviews written in Turkish. A dataset of 3,000 reviews of 10 popular movies was collected and manually labelled. Additionally, another 2,000 synthetic audience reviews containing spoilers were generated with ChatGPT for the same movies. In classification experiments using transformer architecture based deep learning models, tests were carried out first with the original dataset and then the synthetic data were included in the training phase. As a result of the experiments, the highest performance was obtained as 95.09% F1-score with the BERTurk (128K) model in the test where synthetic data was also included. Also, it was found that the inclusion of synthetic data in the training process generally increased the performance.