Filtering Relevant Comments in Social Media Using Deep Learning
David Ramamonjisoa, Hidemaru Ikuma, Riki Murakami · 2022
This paper presents an online tool to filter automatically comments on YouTube based on five models trained on several datasets with a deep learning algorithm. Those datasets were collected from the Kaggle Competition site. We used the language model BERT to realize the comments classification. As BERT is a pre-trained network for masked language modeling or next-sentence prediction, it has to be fine-tuned for document classification. We spent several days fine-tuning our models and testing several techniques to build the last layer of a new model classifier based on BERT. We present in this paper our experiments and results. We obtained a good accuracy on the test set for each model. We used BERT-Tiny (the smallest BERT model) in order to deploy the model on our server and make faster the prediction time for classifying new comments. Our online tool connected in real-time to the YouTube API is ready to serve any queries from users. We describe the web user interface and feedback from more than 50 users during the first six months of the online launch.