A Review of Out-of-Distribution Generalization in Text Classification Models

A basic premise of standard machine learning is that the training set and the test set come from the same distribution, but this premise is often broken in practice. Text classification is one of the most fundamental tasks in natural language processing (NLP). The poor robustness of its models to out-of-distribution (OOD) samples has become an important research topic. This study reviews the latest research on OOD generalization in text classification. First, it defines OOD generalization, which can be viewed from the data perspective as systematic data variance and from the model perspective as distorted features. Second, it introduces domain generalization methods, anti-shortcut feature learning, solutions for class distribution shift, and the main methods for OOD generalization of large language models. Then, it presents evaluation benchmarks and datasets used for three types of evaluation: distribution shift, compositional generalization, and class distribution shift. Finally, it identifies the limitations of existing studies and proposes future research directions.