【摘要】 文本语义相似度的计算在自然语言处理领域的许多应用中发挥着重要作用。BERT等预训练语言模型因为其更优的性能被广泛应用于语义相似度任务中。通过有监督的方法对BERT进行微调，能够得到较高质量的句子表示，但需要大量的已标注数据集。在缺少标注样本的情况下，提出一种无监督的方法，首先，采用对比学习，以无监督的方式训练模型，不依赖于有标签样本。其次，基于提示学习，对输入文本添加固定的模板作为提示，在微调过程中对提示向量进行优化。在SentEval数据集上的实验结果表明，与其他无监督方法相比，该方法有较好的提升。更多还原

【Abstract】 The computation of textual semantic similarity plays an important role in many applications in the field of natural language processing. Pre-trained language models such as BERT are widely used in semantic similarity tasks because of their better performance. Fine-tuning BERT through a supervised approach yields higher-quality sentence representations,but requires a large labeled datasets. In the absence of labeled samples,an unsupervised method is proposed. Firstly,contrastive learning is used to train the model with unsupervised learning,which does not depend on labeled samples. Secondly,based on prompt learning,a fixed template is added to the input text as a prompt,and the prompt vector is optimized during fine-tuning. The experimental results on SetEval dataset show that this method has better improvement compared with other unsupervised methods.更多还原