Transfer learning pro analýzu textových dat

Abstract

The aim of this bachelor thesis was to test transfer learning methods on different datasets and then compare results with simpler machine learning methods. Text analysis is a complex field, so I picked a subfield called text classification. Experiments need data, therefore I included a part dedicated to their preprocessing. There is more than one language used in the experiments. Used languages are English, French and Czech, with more languages I could compare results of each method and model within the language and without the doubt I could tell which method performs the best for the language. I would like to mention a very good performance of the transformer models, they can perform surprisingly well even with small training dataset, in most cases they even outperformed deep learning methods trained on tens of thousands training samples.