The patents of information and communication technology (ICT) in construction are valuable sources of technological solutions to communication problems in the construction practice. However, it is often difficult for practitioners and stakeholders to identify the key communication functionalities from complicated expressions in the patent documents. Addressing such challenges, this study develops a deep learning model to enable automatic recognition of communication-oriented entities (CEs) from patent documents. The proposed model is structured based on the Transformer, consisting of feed-forward and self-attention neural networks to better recognize ambiguous and unknown entities by utilizing contextual information. The validation results showed that the proposed model has superior performance in CE recognition than traditional recurrent neural networks (RNN)-based models, especially in recognizing ambiguous and unknown entities. Moreover, experimental results on some research literature and a real-life project report showed satisfactory performance of the model in CE recognition across different document types.

建筑中的信息和通信技术（ICT）专利是建筑实践中解决通信问题的技术解决方案的宝贵来源。但是，从业人员和利益相关者通常很难从专利文件中的复杂表述中识别出关键的沟通功能。为应对此类挑战，本研究开发了一种深度学习模型，可自动识别专利文件中面向通信的实体（CE）。所提出的模型基于Transformer构建，该Transformer由前馈和自我注意神经网络组成，可以通过利用上下文信息更好地识别模糊和未知的实体。验证结果表明，与基于传统递归神经网络（RNN）的模型相比，所提出的模型在CE识别方面具有更好的性能，尤其是在识别模棱两可和未知实体方面。此外，一些研究文献和真实项目报告的实验结果表明，该模型在不同文档类型的CE识别方面表现令人满意。