Existing automated code checking (ACC) methods require the extraction of requirements from building codes and the representation of these requirements in a computer-processable form. Although these methods have achieved different levels of performance, all of them are still unable to represent all types of building-code requirements. There is, thus, a need to enhance the semantic representations of building codes towards facilitating the representation of all requirements. To address this need, this paper first proposes a new approach to annotate and represent building-code sentences using requirement units that consist of semantic information elements and simple logic operators. To evaluate the proposed building-code annotation approach, this paper also proposes a new natural language generation (NLG)-based method for evaluating annotation quality. The proposed method consists of four steps: data preparation, data preprocessing, NLG model development and training, and sentence evaluation. Sentences from the International Code Council (ICC) building codes were used in the evaluation.