Deep learning models for text classification have been increasingly used in intelligent tutoring systems and educational writing assistants. However, the scarcity of data in many educational settings, as well as certain imbalances in counts among the annotated labels of educational datasets, limits the generalizability and expressiveness of classification models. Recent research positions LLMs as promising solutions to mitigate the data scarcity issues in education. In this paper, we provide a systematic literature review of recent approaches based on LLMs for generating textual data and augmenting training datasets in the broad areas of natural language processing and educational technology research. We analyze how prior works have approached data augmentation and generation across multiple steps of the model training process, and present a taxonomy consisting of a five-stage pipeline. Each stage covers a set of possible options representing decisions in the data augmentation process. We then apply a subset of the identified methods to three educational datasets across different domains and source languages to measure the effectiveness of the suggested augmentation approaches in educational contexts, finding improvements in overall balanced accuracy across all three datasets. Based on our findings, we propose our pipeline as a conceptual framework for future researchers aiming to augment educational datasets for improving classification accuracy1.