The quality of requirements specification directly affects software quality, but quantitatively assessing the quality of requirements specification is a difficult task. In this study, we propose a framework for quantitatively evaluating requirements specifications that improve TextRank and RoBERTa. Aiming at the problem of redundant information in long texts and the difficulty of accurately calculating the similarity, an adaptive TextRank keyword extraction algorithm incorporating dependency parsing is proposed; to deeply excavate the complex structure and deep semantics of the text, a Siamese-RoBERTa model optimizing the attention pooling layer is proposed to capture the important features of the text more comprehensively; finally, the semantic similarity between texts is calculated by cosine similarity to realize the quantitative evaluation of the requirements specification. The experimental results show that the proposed evaluation framework has good performance, its Spearman coefficient reaches 0.973, and the average absolute error between the model evaluation results and the expert evaluation does not exceed 0.5.