Дослідження сучасних методів аугментації текстових даних

Abstract

The goal of this work is to explore methods of text data augmentation, which involves creating new synthetic data similar to real ones, for machine learning tasks where available data is limited. Generative data augmentation is used to combat overfitting, but it has found limited application in Natural Language Processing. Simple augmentation methods like random insertions, replacements, and shuffling are too limited in their effectiveness. Substituting n-grams with synonyms is another method that can be used for data augmentation, as well as the application of intelligent models like Back translation and Style augmentation. The use of generative models such as C-BERT is a popular solution for the augmentation task. Prompt engineering is also becoming increasingly popular for creating queries that prompt the model to provide optimal responses.