Text Classification via Semantic Representation and Similarity

Text classification focuses on categorizing raw texts to specific predefined groups. Text classification consists of intent classification, topic classification, sentiment classification, and language identification due to different classification goals. There are several emerging challenges for text classification applied in real scenarios, including out-of-scope classification, zero-shot classification, few-shot classification, etc. Addressing these challenges requires advancements in model architectures, training techniques, data augmentation strategies, and evaluation metrics. In this thesis, we provide three lines of works to improve the performance of text classification on transformer-based models figuring out one or more challenging problems via semantic representation and similarity.