A Parameter-Independent Fewshot Text Categorization In-corporating Semantic Information

Deep learning is currently widely used in text classification due to its high accuracy. However, deep learning tends to require a large amount of training resources, while deep learning also tends to suffer from overfitting with fewer labeled samples. Jiang et proposed an alternative scheme without training parameters that is simple, lightweight, and versatile in text catego-rization: a simple text compressor (e.g., gzip) combined with a k-nearest-neighbor classifier is used for text categorization. However, npc-gzip only considers the presence of identical words, ignoring semantic similarity. Therefore, we propose a method based on that paper to integrate semantic information, which outperforms the non-pretrained baseline method, npc-gzip, in the few-shot setting.