To solve the problem that construction safety requirement information hidden in project documents is hard to be discovered without relevance and semantic ambiguity, a two-stage integration framework combining NLP techniques was developed for project document analysis and classification and extraction of requirement information. First, the safety targets of the project to be evaluated were obtained by combining the multivariate techniques of NLP, and an association model was established based on the topic model to recommend the appropriate requirement types. Then, the semantic features of the three types of elements were considered, and keyword analysis, sentiment analysis, and dependency analysis were adopted to extract the three types of elements, respectively. Finally, two types of construction projects (civil and industrial) were used as case to test the type recommendation and extraction of construction safety requirements. The results show that the two-stage integration framework recommends four appropriate requirement types for civil and industrial buildings respectively, and the combination of lexical properties and lexical sentiment can effectively extract the requirement keywords and behavior opinion words, and the extraction accuracy of the main elements can reach 88.6% after supplementing the description of building types. The test results confirm that responding to safety target features can recommend suitable types from the complicated requirement information, and the classification and extraction of requirement information combined with NLP avoids subjective preferences and improves the accuracy of information extraction.