A Tibetan Text Classification Method Based on Hybrid Model and Channel Attention Mechanism

Tibetan text classification is a fundamental and essential branch of Tibetan natural language processing. However, due to the scarcity of annotated data, progress in related research has been slow. To alleviate the problem of scarcity of Tibetan text classification data and poor classification effect, we first crawl and pre-process 1,762 Tibetan text classification data from websites such as China’s Tibetannet and TIBET.CN and integrate these data into the open-source TNCC (Tibetan News Classification Corpus) dataset. Then, based on this augmented dataset, we propose a classification approach using the hybrid model and channel attention mechanism, including the following models: CINO+CNN+BiLSTM+SE and CINO+CNN+BiLSTM+ECA. Finally, five different models are used as baselines, including CINO, CINO+CNN, CINO+BiLSTM, CINO+BiLSTM+Att, and CINO+CNN+BiLSTM+Att. The experimental results show that the models proposed in this paper can effectively leverage the advantages of the hybrid model and weighting method of channel attention. CINO+CNN+BiLSTM+ECA achieves the best performance, with an F1 score of 76.4 %.