Text mining technology enables the processing and analysis of large amounts of text data to improve the efficiency of linguistic analysis and information processing, providing support for a variety of aspects ranging from architectural design to sustainability assessment. In this paper, we propose a corpus extraction method based on BiLSTM (Bidirectional Long Short-Term Memory Network) to train a natural dimension evaluation index system by extracting information highly related to the natural dimensions of buildings from massive architectural text data. The active learning strategy is also applied in to improve the accuracy and efficiency of BiLSTM. Based on the constructed corpus, the natural dimensions of building design are experimentally categorized into 9 major evaluation indexes, such as light, water, wind, natural environment, indoor and outdoor space, building structure, landscape view, material, orientation.