Few-shot English text classification method based on graph convolutional network and prompt learning

Abstract

The classification method based on pre-training fine-tuning usually requires a large amount of labeled data, which makes it impossible to apply to few-shot classification tasks. Therefore, this paper proposes a novel few-shot English text classification method based on graph neural network and prompt learning. The text level graph convolutional network is used to construct a graph for each input text and share global parameters, and the result of the text graph neural network is used as the input of the prototype network. This new method generates a class representation vector rich in class semantic information for each class. On the other hand, a manual prompt template is used to obtain a class prediction semantic vector for the [MASK] position. In the process of classification prediction, the similarity between class prediction semantic vector and class representation vector is used as classification basis. Compared with the traditional method of using linear layer for final answer mapping and the method of using custom class representation word set for classification prediction, this new method alleviates the semantic loss in the process of answer mapping. Through random sampling on the three data sets of THUCNews, SHNews and Toutiao, a few-shot training set and a verification set are formed for the experiment. The experimental results show that the proposed method has improved the overall performance of the 1-shot, 5-shot, 10-shot and 20-shot tasks on the above dataset, especially the 1-shot task. Compared with the baseline few-shot text classification method, The accuracy is improved by 7.59%, 2.11% and 3.10% respectively, which verifies the effectiveness of the proposed method in few-shot English text classification.