Effect of training fragment length on Transformers in text complexity prediction

Rafik Hachana, Vladimir V. Ivanov

Keywords: Text complexity classification, machine learning, transformers, bert

TL;DR: We studied the effects of using different text fragment lengthes on the performance and training time for models in the text complexity classfication task

Abstract: With the myriad practical applications of text complexity classification, it is important to optimize the training text fragment size for performance. We experiment with fine-tuning pre-trained BERT models to classify the complexity of Russian school text using different fragment sizes for training.