【摘要】 问题生成(QG)是自然语言处理中一个具有挑战性的任务，其目标是根据不同类型的数据，生成语法正确且语义相关的问题。目前，融合答案信息的问题生成方法主要采用序列到序列的神经网络模型，但这些方法存在以下问题：(1)对RNN模型的依赖性高；(2)欠缺捕捉输入文本的语义信息；(3)缺乏对少数民族语言中问题生成的研究。针对以上问题，该文通过一种基于答案感知的BERT自回归方法改进了藏文问题生成。首先，该方法利用藏文预训练模型BERT来处理问题生成任务；其次，通过重组输入部分以进一步提升问题生成的性能，即不断将新生成的词元追加到输入文本中，直到预测到特定的结束标记，使其变为一种连续的生成方式，从而改善了生成的连贯性；最后，为了增强问题和答案的关联，该文通过标记答案位置的方式来指示问题生成，以消除歧义并提高问题的质量。经过实验验证，该文所使用的方法在藏文问题生成任务中表现出明显的性能提升，相较于基线系统，生成的问题更准确和更连贯。更多还原

【Abstract】 Question Generation(QG) is a challenging task in natural language processing, with the goal of generating grammatically correct and semantically related problems based on different types of data. This paper improves Tibetan question generation through a BERT autoregressive method based on answer perception. Firstly, this paper utilizes the Tibetan pre-trained model BERT to handle question generation tasks. Secondly, by restructuring the input part to further improve the performance of question generation, the newly generated token is continuously added to the input text until a specific end tag is predicted. Finally, to enhance the correlation between the question and the answer, the paper introduces the position of the answer to eliminate ambiguity and improve the quality of the question. Experiments show that the proposed method brings significant performance improvement in Tibetan question generation tasks.更多还原