FHTC: Few-Shot Hierarchical Text Classification in Financial Domain

As an extensively applied task in the domain of natural language processing, text classification has moved a long way since deep learning technology develop rapidly. Especially after the pre-trained models arrived, the classification performance has been tremendous improved. However, complicated financial text often has multiple structured labels, and there are also many difficulties to have large amounts of labeled samples to ensure high-quality predictions. The existing competitive classification models can only solve one of the problems. To address these issues, we propose a hierarchical classification structure with two level. In the first level, the basic classifier is enhanced by label confusion algorithm to mine the dependency between labels and samples. In the second level, a few-shot classification model under meta-learning framework can complete the classification task based on the predictions from the previous level and a few labeled training samples. We explain our model on two large Chinese financial datasets, and find that it has superiority in both performance and computational expenditure compared to existing competitive classification model, few-sample classification model and hierarchical classification model.