为了更好地学习藏文语言特征以及探究藏文预训练语言模型的最佳注意力机制头数，将词性与藏文预训练模型相结合，并进行了对比实验确定最佳的注意力头数，旨在提高语言模型对藏文语言特征的理解以及下游任务的性能。实验结果表明，在多个分类任务中，注意力头数为12的预训练模型皆表现了良好的性能。此外，将词性融入预训练模型后，文本、标题和情感分类任务的模型F₁值分别提高了0.57%、0.92%和1.01%。实验结果证明融入词性特征后，模型可以更准确地理解藏文语言结构和语法规则，从而提高分类任务的准确率。

In order to acquire superior Tibetan characteristics and enhance the model's understanding of Tibetan features, part-of-speech was combined with the Tibetan pre-trained language model. Meanwhile, improving the performance of downstream tasks, the optimal attention mechanism head number of Tibetan pre-trained language model were explored by comparative experiments. The results show that pre-trained language models with 12 attention heads perform well in multiple classification tasks. Furthermore, after incorporating part-of-speech into the pre-trained language models, the macroF1 values of text, title and sentiment classification tasks increase by 0.57%, 0.92% and 1.01% respectively. It is conclued that after incorporating part-of-speech features, the language structure and grammar rules of Tibetan can be better understanded.