Abstract: Traditional graph convolutional network methods can effectively model graph structures under data-limited conditions.However,due to their reliance on sparse one-hot encoding,they face limitations in capturing the contextual relationships between words.This issue is particularly pronounced in low-resource language environments.Taking the Sawcuengh language text topic classification task as an example,this task faces not only data scarcity but also the challenge of complex linguistic structures.To address these challenges,this paper proposes a Sawcuengh language topic classification method suitable for low-resource settings－ZHA_TGCN.This method leverages the Sawcuengh pre-trained model,ZHA_BERT,to extract textual features,and combines these features with Sawcuengh tone features.These combined features are then input into a BiGRU to learn deep semantic representations.The learned representation vectors are used as node features for the GCN,which propagates labels to learn the feature representations of both the training data and the unlabeled test data.Finally,a Softmax layer is used to output the classification results.Experimental results show that the proposed method achieves an accuracy of 82.12%,precision of 90.08%,recall of 92.46%,and an F1 score of 90.18% in the low-resource Sawcuengh language topic classification task,demonstrating the effectiveness of the method.

Key words: Low-resource language, Sawcuengh language, Subjet classification, Pre-trained model, Graph convolutional network