基于LSTM的《红楼梦》文本风格分界点识别方法

A method of identifying the style demarcation point of A Dream of Red Mansions based on deep learning

针对《红楼梦》文本风格分析中,手工设计的特征带有主观性与仅根据单个分界点前后的差异直接判断分界点位置的问题,提出了一种基于深度学习的文本风格分界点识别方法及用于文本分类的深度模型1-DconvLSTM。该方法在不同的训练分界点处,相互独立的训练多个结构相同的1-DconvLSTM模型。通过测试模型的泛化性能,得出在不同训练分界点处前后差异显著性的大小,并据此确定文本风格的实际分界点位置。实验结果表明,红楼梦的文本风格分界点在第80回,强有力的支持了《红楼梦》80回前后不是同一作者所著的结论。

Aiming at the problems of manual design features in traditional A Dream of the Red Mansions text style analysis with subjective issues and directly judging the position of the style demarcation point based on the difference between the single demarcation point before and after,this paper proposes a text demarcation point recognition method based on deep learning and a text classification model 1-DconvLSTM.First,the method trains multiple 1-DconvLSTM models with the same structure independently at different training cut-off points,and then tests the generalization performance of these models to obtain the magnitude of the significant difference between the different training cut-off points.Based on this,the actual demarcation position of the text style is determined.The results of the experiments show that the text style demarcation point of A Dream of the Red Mansions is at 80 chapters,which strongly supports the conclusion that A Dream of the Red Mansions is not written by the same author around 80 chapters.