关键词: 语义相似度, 预训练模型, 隐含Dirichlet分布,  , 新闻评论

Abstract: Aiming at the problem that the pre-training model would cut off part of text when dealing with long text such as news, which led to the loss of text infomation, we  proposed a  method to build a model by combining  TextRank algorithm, implicit Dirichlet distribution topic model and pre-training model on the basis of integrating news title information, and  compared the model with other  semantic similarity calculation methods. The results show that the accuracy rate of the model is 82.46%, the recall rate is 87.43%, the accuracy rate is 82.68%, and the F₁ value is 84.99%,  the optimal results are obtained, which effectively improves the performance of semantic similarity calculation between news texts and comments.

Key words: semantic similarity, pre-training model, implicit Dirichlet distribution,  , news comment