COMPARATIV ANALYSIS OF TEXTUAL MODELS BERT, BART AND XLNET FOR PREDICTING THE POPULARITY OF YOUTUBE VIDEOS

Abstract

This paper talks about a comparative analysis of three text models: BERT, BART, and XLNet. The data for training and validation were taken from the website kaggle.com. There are two datasets: the initial one and an expanded one, created in an attempt to improve the recall metric. After downloading, preprocessing was done, including tokenization and stemming. The dataset was split in a ratio of 80:20 for training and validation. Different values for learning rate and batch size were tested, and the best result was achieved using the BERT large model with a learning rate of 1e-5 and a batch size of 8. The other two models produced somewhat worse results than BERT. All experiments and results will be presented below.