Part-of-speech tagging of building codes empowered by deep learning and transformational rules

Abstract

Automated building code compliance checking systems were under development for many years. However, the excessive amount of human inputs needed to convert building codes from natural language to computer understandable formats severely limited their range of applicable code requirements. To address that, automated code compliance checking systems need to enable an automated regulatory rule conversion. Accurate Part-of-Speech (POS) tagging of building code texts is crucial to this conversion. Previous experiments showed that the state-of-the-art generic POS taggers do not perform well on building codes. In view of that, the authors are proposing a new POS tagger tailored to building codes. It utilizes deep learning neural network model and error-driven transformational rules. The neural network model contains a pre-trained model and one or more trainable neural layers. The pre-trained model was fine-tuned on Part-of-Speech Tagged Building Codes (PTBC), a POS tagged building codes dataset. The fine-tuning of pre-trained model allows the proposed POS tagger to reach high precision with a small amount of available training data. Error-driven transformational rules were used to boost performance further by fixing errors made by the neural network model in the tagged building code. Through experimental testing, the authors found a well-performing POS tagger for building codes that had one bi-directional LSTM trainable layer, utilized BERT_Cased_Base pre-trained model and was trained 50 epochs. This model reached a 91.89% precision without error-driven transformational rules and a 95.11% precision with error-driven transformational rules, which outperformed the 89.82% precision achieved by the state-of-the-art POS taggers.

Introduction

Efforts to automate code compliance checking started more than half a century ago when Fenves (1966) developed decision tables to automatically check the design of steel structures [1]. The success of compliance checking decision table inspired more researches in this area. Examples include a computer-aided design (CAD) system for 2D and 3D steel structure called STEEL-3D [2], an expert system for reinforcement concrete design [3], a rule-based application for structure members [4], and a knowledge-based system for multiple building codes [5]. More advanced code compliance checking software was then developed. The Construction and Real Estate Network (CORENET) by Singapore Building Construction Authority was capable of checking 3D industry foundation classes (IFC) data model [6]. The Express Data Manager (EDM) Suite by Jotne EPM Technology allowed code checking on Building Information Modeling (BIM) data [7]. The BCAider by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) in Australia enabled automatic compliance checking against Building Code of Australia (BCA) [8]. The Solibri Model Checker (SMC), a BIM-powered automated code compliance checking system by Solibri, achieved rule-based code compliance checking by user-customized plugins [9]. Patlakas et al. developed a BIM-based system to check code compliance of timber structure design automatically [10]. Fang et al. developed a deep learning-based method to automatically check if a site worker complies to code of their certification [11]. The combination of BIM and automated code compliance checking systems increases the theoretical benefit of BIM in the construction industry. However, according to a survey by Smits et al. (2017), the actual benefit of implementing BIM in construction projects is still limited [12]. The authors suggest that the narrow range of checkable codes of most recent automated code compliance checking tools may limit the actual benefit of BIM. Even for the narrow range of checkable codes, they are usually oversimplified. The oversimplified codes are not enough to support the increased project complexity and creativity of designers and, therefore, could negatively affect the benefit of adopting BIM for users and owners [13].

The narrow range of checkable codes also limit wide applications of these automated code compliance checking systems. Extending the range of checkable building code requirements emerges as an urgent need in the development of automated code compliance checking systems. Natural Language Processing (NLP) powered by Part-of-Speech (POS) tagging has been proposed to automate the building code requirements extraction [30] and, therefore, extended the range of checkable building codes of automated code compliance checking systems and reduced the needed manual efforts in such extraction [14], [15], [16]. NLP and deep learning have many applications in the Architecture, Engineering, and Construction industry (AEC) [21]. For example, Fang et al. developed a text classification method with deep learning to spot near misses in safety reports [17]. Zhong et al. used a deep learning method to classify building quality problems [18]. Trappey et al. used attention mechanism to generate summaries of engineering patents [19]. High performance was achieved but POS tagging error was identified as one major source of error of the whole system. Accurately POS-tagged building codes are desired to support such NLP-based automated building code compliance checking. Existing generic POS taggers, however, cannot provide such high accuracy on processing building codes [20].

The authors are therefore proposing a new POS tagger that is tailored to building codes. The intent of the study is to improve the accuracy of POS tagging on building codes. Accurate POS tagging results are needed to support successful code requirements processing for accurate automated code compliance checking. The proposed POS tagger combines neural network model and error-driven transformational rules. Neural network model and error-driven transformational rules together make the proposed POS tagger outperformed the state of the art. The proposed POS tagger reached a 95.11% accuracy, which is higher than the 89.82% accuracy achieved by the state of the art.

In practice, this POS tagger plays an important role in those NLP-based automated code compliance checking system frameworks similar to [14] (Fig. 1), and in NLP-based automation systems in the AEC domain in general [22]. This research can boost the accuracy of the POS tagging therefore support automated building code compliance checking systems and NLP-based systems in the AEC domain. Accurate POS tagging results of building codes is vital to a high performance of the extraction of engineering knowledge embedded in the building codes. The background automated code compliance checking system framework in Fig. 1 contains an automated regulatory information extraction component (which uses a POS tagger) that converts building code requirements to logic clauses, an automated building design information extraction component that extracts building design information from Building Information Models (BIMs), and an automated reasoning component that outputs the code compliance report. The automated regulatory information extraction component can use the proposed POS tagger, which is illustrated in Fig. 3. This system is fully automated from the end-user’s perspective. The automated building code compliance checking system takes a rule-based approach to extract information from building codes automatically. Although the POS tagger uses neural network model which is probabilistic in training, the developed POS tagger as a result of the training is deterministic. The weights of the neural network are fixed after the training, leading to determinist results when applying the POS tagger. Therefore, with a robust POS tagger and other well-performing components, the NLP-based automated building code compliance checking system has a better chance to detect all noncompliance cases in a building design without intervention from the user. Due to the imperfect (i.e., less than 100%) precision and recall in the state-of-the-art NLP-based building code compliance checking systems, some manual intervention will still be needed to fix errors in the extraction results of the embedded engineering knowledge in the building codes. Such manual intervention is expected from the developers, not from end users. In addition, the amount of manual efforts needed to fix automatic extraction errors is minor comparing to those needed in a manual extraction. In this paper, the authors propose to boost the performance of NLP-based automated code compliance checking systems by providing more accurate POS tagging results to such systems.

The remainder of this paper is organized as follows. Section 2 explains the technical details of part-of-speech tagging, error-driven transformational rules, recurrent neural network, and computing techniques to avoid overfitting, used in this research. Section 3 describes the proposed POS tagger. Section 4 presents the experiment to test the performance of the proposed POS tagger. Section 5 illustrates and discusses the results of the experiment. Finally, 6 Contributions to the body of knowledge, 7 Limitations and future work, 8 Conclusion present the limitations and future work, contributions to the body of knowledge, and conclusion of this research, respectively.