A deep natural language processing‐based method for ontology learning of project‐specific properties from building information models

Element property is a crucial aspect of building information modeling (BIM) for almost all BIM‐based engineering tasks. Since there are limited properties predefined in Industry Foundation Classes (IFC) specifications, a vast number of property concepts were customized and stored in BIM models, which lack labor‐intensive data modeling and alignment for effective information management and reuse. To tackle the challenge, this study presents a natural language understanding (NLU)‐based method for the automatic ontological knowledge modeling of project‐specific property concepts from BIM models. A soft pattern matching model was used to acquire contextual definitions of concepts from a domain corpus before applying deep NLU models to transform the concept names and definitions into dense vector representations. These outputs were then fed into two stacking ensemble learning models to carry out two tasks: (a) classifying whether an unseen concept overlaps with the IFC ontology, and (b) aligning the repetitive concepts with the most relevant concepts in the ontology. Finally, all fresh properties were appended to an IFC ontology, either as new objects or new synonyms. The performance was evaluated based on 327 property concepts from real‐life BIM models. The results show that the proposed approach incorporating reading comprehension of definitions outperforms the existing name similarity‐based methods. Finally, a case study on a renovation project demonstrates the effectiveness of this study in automatic ontology modeling of property concepts.

ferent types of information (Eastman et al., 2011).These semantically rich building models can be applied in a range of engineering applications that add practical values to architecture, engineering, and construction (AEC) projects (Xue et al., 2018), such as clash detection (Hu et al., 2019), progress monitoring (Golparvar-Fard et al., 2015), and maintenance management (Chen et al., 2022).Moreover, based on the Industry Foundation Classes (IFC) standards (buildingSmart International Ltd., 2019), project information and knowledge can be stored in vendor-neutral IFC file formats, which allows BIM information to be flexibly exchanged between stakeholders and various engineering software.
Construction projects are highly complex, projectdependent, and one-off productions involving multiple stakeholders (Karim & Adeli, 1999a, 1999b).Since each project operates within an exclusive spatio-temporal context with different designs and engineering solutions (Adeli & Yeh, 1989;Zhang et al., 2022), the information requirements and naming applied to the BIM models can vary (Chen et al., 2017).Although IFC standards are continuously updated with increasing numbers of domain concepts and terminologies, they are still insufficient to cover the knowledge body and associated concept names in certain areas.
The element properties and quantities are two of the most typical kinds of BIM information that lack complete data modeling in IFC specifications.A property (IfcProperty) describes a particular semantic characteristic of an object (e.g., "FireRating") from a professional viewpoint, such as structural properties or environmental impacts.At the same time, a quantity (IfcPhysicalQuantity) refers to a quantitative measure of an element, such as geometry and weight (buildingSmart International Ltd., 2019).The latest IFC standards (IFC 4 ADD2 TC1; buildingSmart International Ltd., 2017) have 420 property sets and 93 quantity sets containing predefined property and quantity concepts associated with various entity types.For example, the IfcWall has 115 applied properties distributed across 12 property sets (e.g., Pset_WallCommon).However, those predefined properties and quantities are still insufficient for real-world BIM projects.Consider that the beam elements, the seat depth, web thickness, and chord length are essential data requirements to understand the structural information of beams in a BIM model, but none of them is covered by the predefined properties in the current IFC specifications.
To cope with issues where project information exceeds the scope of the data model, IFC provides mechanisms to allow users to define customized property and quantity concepts to represent different information and knowledge of objects and buildings in BIM projects (Hu et al., 2021).Herein, project-specific properties (PPs) are used to represent these customized property and quantity instances (e.g., "room bounding") that are not within the built-in property set definition (PSD) and quantity set definition (QTO) of IFC, in contrast to standard properties (SPs) that stand for predefined properties (e.g., "IsExternal") and quantities in IFC specifications.
Based on this extension mechanism, various PPs can be efficiently created by stakeholders and used within specific projects.However, these PPs typically lack standardization and integration with general standards, limiting data reusability in construction projects.Hence, as the IFC specifications still have rooms for further enrichment of data schema, it is beneficial to iteratively standardize and integrate these PPs into the IFC data model to achieve efficient information exchange and management.Nonetheless, manually collecting the PPs in BIM models and conducting information modeling of new concepts is a slow, cumbersome, and labor-intensive process because of the vast number of domain-specific concepts.
This problem can be viewed as one of ontological knowledge modeling.A philosophical concept, ontology, is the study of being or existence (El-Gohary & El-Diraby, 2010;Zender & Humm, 2022).Recently, ontologies have been leveraged to represent IFC data models based on semantic web technologies (Pauwels et al., 2017), such as the Web Ontology Language (OWL; W3C, 2013).Despite efficient tools (e.g., standard query languages) in the knowledge engineering domain that can now be applied to BIM data processing, acquiring domain knowledge and constructing ontologies manually is very time-consuming and errorprone (Petasis et al., 2011).Likewise, there is a heavy workload for ontological knowledge modeling of massive PP concepts regarding the IFC ontology.
There is a strong a priori case for tackling these issues by applying ontology learning (OL) techniques to acquire the PP concepts and enrich the IFC ontology automatically.In the knowledge engineering domain, OL is a field that automatically constructs, enriches, and maintains ontologies (Maedche & Staab, 2004).The existing OL systems mainly rely on natural language processing (NLP) to extract ontological knowledge (e.g., concepts, synonyms, and axioms) from textual documents (Konys, 2018) because they are primary carriers of domain knowledge.Nonetheless, the situation differs when it comes to ontology modeling of property concepts for the AEC domain because many element properties have already been defined and isolated as instance entities in BIM models to reflect information and knowledge in professional processes.So, it is important to build a new OL system that can learn PP entities from structured BIM models.
The main challenges arise from concept identification.During the OL process, some extracted PPs are new concepts, whereas others overlap with the existing ontology.However, it is difficult to distinguish them because a property concept can have very dissimilar name descriptions across different projects.Besides, some disparate property concepts could have names that are semantically close to each other (e.g., "FrontSurfaceArea" [PP] and "Out-terSurfaceArea" [SP]).It is unreliable to use the widely adopted name similarity (NS) measure methods (e.g., word embeddings) to predict the type of PPs, which can make the resulting ontology mis-organized, redundant, and inconsistent.
To tackle these challenges, this study proposes a novel natural language understanding (NLU)-based OL approach that automatically assimilates property information stored in BIM models and enriches the ontological knowledge models.NLU is a sub-branch of NLP that deals with the reading comprehension of texts (Allen, 1988).It is based on statistical machine learning models that are widely used in the construction field (Rafiei & Adeli, 2018;Rafiei et al., 2017).In the proposed method, the definitions of PPs are automatically extracted from the textual corpus and read by the deep NLU models to predict their relationships with the IFC ontology and determine how the ontologies are edited.The expected outcomes are (a) an enriched IFC ontology with new property concepts and synonyms and (b) updated queries, inference rules, and BIM models that allow better retrieval and reasoning of building data.This method accelerates the ontology construction process so as to broaden the coverage of BIM ontologies.A wider range of standardized property concepts can be utilized and exchanged, therefore improving a BIM system's accuracy, usability, and use value.
The remainder of this paper is structured as follows.Section 2 introduces the background of this study; Section 3 presents the proposed method; Section 4 illustrates the experiment design and performance evaluation; Section 5 introduces a case study on a realistic project; Section 6 discusses the study's contributions, implications, and limitations; and Section 7 concludes by outlining the significance of this research.For convenience, the list of abbreviations used in this paper is presented in Table 1.

2.1.1
Data exchange of object properties and property sets based on IFC IFC is a widely used open-source data standard for explaining, exchanging, and sharing BIM data among different software applications in the AEC industry (ISO, 2018).The IFC data schema is defined in the EXPRESS data modeling language (ISO, 1994(ISO, , 2018)), which contains concepts, terms, and data specification items derived from the use within trades, disciplines, and professions A property in IFC refers to "an abstract generalization for all types of properties associated with IFC objects through the property set mechanism" (buildingSmart International Ltd., 2019).The data type of an individual property can be a single value, an enumerated value, a reference value, or a table value.A property set (IfcProp-ertySet or Pset) performs as a container that holds related properties within a property tree (buildingSmart International Ltd., 2021b).Some Psets are predefined as PSD in the IFC specifications and assigned to certain object occurrences and types.Similarly, IfcQuantitySet is a container that holds the individual quantities (buildingSmart International Ltd., 2021c) and some quantity sets are predefined as QTO in IFC specifications.A quantity is defined as "the measurement based on scope metric (e.g., area, length, count, etc.)" (buildingSmart International Ltd., 2021a).
The predefined property and quantity concepts have well-documented definitions in PSD and QTO for standard exchanges.However, because of the limited coverage of default parameters (Zhang et al., 2014), many studies have extended the Psets and properties within the IFC schema to support different applications (Cemesova et al., 2015;Motamedi et al., 2016;Vectorworks, 2018).For example, Maltese et al. (2017) created a custom property set called Pset_RatingResult, which was associated with IfcBuilding, to store and manage rating system data for sustainable building assessments.
Many software applications, such as ArchiCAD, Bentley Systems, and Revit, support IFC-based data exchange in the construction industry.The conversion of internal data models of various software to an IFC data file needs to map different levels of information related to element geometry and semantics.Even though object types and geometries can be accurately converted into IFCcompliant models in most systems, mapping built-in object properties with IFC-predefined properties and property sets is often inadequate (Ying et al., 2022).For example, Autodesk Revit allows some basic parameters, such as "level" or "height," automatically mapped to corresponding IFC properties.However, many properties still need to be manually aligned with SPs using parameter mapping tables (Autodesk, 2018).Any object properties not explicitly defined by the IFC specifications are placed in the custom Psets during the model conversion.Various commercial BIM software applications adopt a similar mechanism to support users in defining and exchanging the custom properties with the IFC data model.
In general, PPs arise from the following aspects of BIM projects.First, since different software platforms are used in various projects, the property concepts in the software data models could cause a mismatch with the IFC data model (Lee et al., 2018;Mirarchi & Pavan, 2019).Some concepts are not covered by the built-in PSD and QTO, while others are covered but lack alignment by software vendors.Second, many property concepts that originated from the different professional requirements of stakeholders could be customized, and the relevant data could be generated in BIM software applications.These property concepts are often placed in custom property sets in IFC during the conversion process because mapping parameters by endusers could be challenging and problematic.Given that there are hundreds of predefined Psets and thousands of SPs in the IFC standards, a precise alignment requires an in-depth understanding of the meanings of all SPs and a laborious identification of similar ones from the candidate concepts.

2.1.2
Ontology-based representation of IFC-based BIM models Despite the widespread adoption of the IFC data schema in the construction industry, the use of the Standard for The Exchange of Product model data limits the expressiveness of IFC languages since formal semantics and contexts in the content are not provided (Barbau et al., 2012).This causes difficulties when integrating IFC with other heterogeneous data sources, as well as the combination of external ontological resources and the reuse of engineering ontologies (Beetz et al., 2009).
To address these issues, many studies have attempted to convert the IFC schema into equivalent OWL ontologies (Beetz et al., 2009;Hoang, 2015;Terkaj & Šojić, 2015).Among these works, the ifcOWL ontology developed by Pauwels and Terkaj (2016) has been officially adopted by buildingSMART (2021a).The ifcOWL ontology is formulated in the OWL2 Description Logic (DL), and its structure closely aligns with the original EXPRESS schema.
Due to the complete transformation of the EXPRESS constructs, the ifcOWL ontology is hampered by its large size and cumbersome instance files.To resolve this issue, the Linked Building Data ontologies proposed by the World Wide Web Consortium provide a solution centered on a lightweight and scalable Building Topology Ontology (Janowicz et al., 2019), where different modular ontologies, such as the Ontology for Managing Geometry (Bonduel et al., 2019), File Ontology for Geometry formats (Bonduel et al., 2019), and Damage Topology Ontology (Wagner & Rüppel, 2019), can be incorporated to describe various building-related information.

2.2
Automatic ontology construction and enrichment

Fundamentals of ontology learning
The construction of ontological knowledge models will benefit from (semi-)automatic OL methods to reduce labor-intensive knowledge engineering works.OL includes the tasks of (a) ontology enrichment, which expands an existing ontology with additional concepts and semantic relationships; (b) inconsistency resolution, which removes the inconsistencies within an ontology; and (c) ontology population, which appends new instances of concepts to the ontology (Petasis et al., 2011).As illustrated by Browarnik and Maimon (2015), an OL process can be deconstructed into six layers increasing in complexity from bottom to top, forming a "layer cake" as shown in Figure 1.Identifying objects and their alternative F I G U R E 1 Ontology learning (OL) "layer cake."Note that the object or term refer to the instances of concepts and relations.
realizations/synonyms from multimedia resources (e.g., text or images) is a fundamental step, and associating them to concepts is known as ontology population.More complex tasks include identifying new class concepts, taxonomy construction, relation extraction, and rule acquisition, all of which belong to ontology enrichment.Additionally, inconsistency checking is operated throughout the OL process to ensure the accuracy of the updated ontologies.
To the authors' knowledge, studies on automatic ontology construction and enrichment in the AEC industry are scarce.Xu and Cai (2021) developed an ontology for the utility infrastructure domain by coupling the semantics of the CityGML Utility Network application domain extension.A base ontology is first generated from Unified Modeling Language data models before an NLP-based approach is applied to learn semantics from the glossaries and enrich the ontology automatically.Pan et al. (2021) proposed a new framework called "video2entities" that extracts entities from videos to update AEC knowledge graphs, which can be deemed as the data layer of the domain ontology.Wu et al. (2022) propose a rule-based information extraction approach for the mechanical, electrical, and plumbing (MEP) semantic web.Named entities can be recognized in the NL texts, and relationships can be extracted using a dependency-path-based matching algorithm.

2.2.2
Automatic concept mapping methods in the construction domain Since the construction industry uses different terminologies, taxonomies, dictionaries, and ontologies to represent knowledge and information (Le & Jeong, 2017), the task of aligning various semantic models to the IFC ontology for BIM-based applications is a complex one that often relies on manual manipulation (Herrera-Martín et al., 2022;Jiang et al., 2022).For example, Tang et al. ( 2020) incorporate the Building Automation and Control Networks (BACnet) protocol into the IFC data model to exchange building automation system-related information.When mapping the BACnet protocol to the IFC, a manual comparison and selection of concepts between the two data models are needed.This manual matching process is quite daunting and laborious, as the heterogeneous sources are often large and complex.Therefore, some studies have investigated how to automatically match concepts in different semantic models.Pan et al. (2008) introduced three semi-automated approaches to map ontological standards that describe the semantics of building models.These methods are based on co-occurrence frequencies in corpora, names, and attribute values to discover related concepts in two heterogeneous ontologies (i.e., The CIMSteel Integration Standards (CIS/2) and IFC).Zhang and El-Gohary (2016) developed methods based on WordNet (Fellbaum, 2012) to match concepts in building regulations to the related IFC concepts and identify their relationships.Recent studies tend to use the advanced word embeddings to measure the semantic similarities (SSs) between concept names.Le and Jeong (2017) used Word2Vec (Peters et al., 2018) to transform words into high-dimensional vector spaces to group concepts in the transportation domain corpus.Zhang and El-Gohary (2020) used word embeddings to represent regulatory concepts and match IFC entities.Zhou and El-Gohary (2021) proposed a method that automatically aligns the semantic information of the BIM models to computerinterpretable regulations based on an ontology and Word2Vec.

Summary
Because of disparate project requirements and processes (Florez-Perez et al., 2022), massive PPs stored in BIM models are beyond the IFC specifications' scope.Automatic ontology modeling of this PP information must be realized to facilitate effective information management, retrieval, and reuse with BIM.Nevertheless, automatic knowledge acquisition is a complex task that combines knowledge representation, logic, philosophy, databases, machine learning, and NLP (Petasis et al., 2011).The current studies on automatic ontology construction and enrichment in the AEC domain deal with tasks in specific subdomains (e.g., utility infrastructure- Xu &Cai, 2021, andMEP-Wu et al., 2022).However, they do not have sufficient generality for BIM ontological modeling.Moreover, existing works mainly focus on developing domain-specific information extraction techniques, while an automatic assessment of the consistency of the resulting ontologies with newly added concepts is rarely discussed.In essence, consistency maintenance of BIM ontologies is essential as an ontology that contains conflicting information is of little use (Petasis et al., 2011).Inconsistency problem induced by duplicate concepts must be tackled in this study because abundant repetitive concepts are found throughout the process of BIM data interoperability.Thus, semantic comparison of newly extracted concepts with the existing concepts in ontologies is necessary.The existing methods that match external concepts with BIM/IFC concepts are based on NS.However, NS-based methods are not suitable for the property concepts because concepts targeting the same semantics often have varied multiword expressions (MWEs).For example, a PP named "embodied carbon" for the IfcWall entity means the total greenhouse gas emissions generated to produce a built asset.Its equivalent SP in IFC is "ClimateChange," which is defined as "the quantity of greenhouse gases emitted calculated in equivalent CO 2 " (buildingSmart International Ltd., 2019).To precisely encode the newly extracted concepts to an ontology, their contextual definitions should be incorporated to compare with the IFC concepts.However, it is unknown how to acquire definitions of property concepts when they are not attached in BIM files and how to make the machine understand the definitions to determine the strategies of ontology editing.
Therefore, this research aims to develop an OL approach that automatically aligns custom properties in BIM models with IFC SPs and enriches the ontology with new objects and synonyms based on the reading comprehension of definitions by machine learning models.The objectives are to automatically (a) extract property concepts from BIM models and search their textual definitions from the domain corpus; (b) determine whether newly extracted concepts are fresh concepts or repetitive concepts to the BIM ontology, and subsequently align them with the most relevant concepts, based on NLU of definitions; and (c) enrich the BIM ontology with new property concepts and synonyms.

THE PROPOSED METHOD
The proposed NLU-based OL method to automatically enrich IFC ontology from BIM instance models is presented in Figure 2. It begins with extracting PPs from BIM models, while the second stage is the key part of the proposed approach.First, the contextual definitions of property concepts concerning the associated objects are collected.If the contextual definitions cannot be found in project-related documents (e.g., guidelines or specifications), they are automatically searched in a corpus that accumulates abundant domain knowledge and information from online resources.Second, the names and accompanying definitions of PPs and SPs are transformed into high-dimensional vector space using a pre-trained word embedding model and a fine-tuned sentence embedding model.Third, the cosine similarity of vector spaces between PPs and SPs is calculated.
The results of both models are taken as the features of two ensemble machine learning models, which in turn conduct containment classification and PP-SP alignments.The containment classification checks whether the target PPs overlap with the current IFC ontology to ensure the existing constructs do not contradict the newly added concepts.When a property concept is identified as belonging to a predefined Pset of the IFC schema, it is matched to the SP ranked first regarding the overall SS score produced by a regressor.In the final stage, a modular IFC Natural Language Expression (INLE) ontology is populated with new property instances and synonyms, and the related queries and inference rules are extended with the ontology evolution.Meanwhile, the name and data type of PP are replaced with the matched SP to improve the semantic interoperability and retrievability of the BIM models.

Overview of INLE ontological structure
Inspired by the concept of Linked Building Data, a modular INLE ontology was developed in the authors' previous work (Yin et al., 2022) using the NL expressions of IFC concepts to achieve NL-based BIM data retrieval.The INLE ontology consists of NLName classes describing the NL names of IFC entities (e.g., NLName_IfcWall) and following the same class hierarchy of ifcOWL ontology (e.g., NLName_IfcWall is a subclass of NLName_IfcBuildingElement).NLName instances represent the realizations of (a) NL names of classes (IFC entities), such as "girder" and "beam" for IfcBeam; and (b) NL names of object instances, such as "concrete" for an IfcMaterial instance.To this end, hundreds of NLName In this study, the INLE ontology is employed as the seed IFC ontology for enriching new property concepts and synonyms due to its appropriate functionality and ontological structure.Based on the INLE ontology, the PPs extracted from BIM models are first checked to determine whether they are identical to any forms of name instances of common properties.Any overlapping concepts are directly aligned with the SPs, and the following OL processes handle the non-overlapping concepts.

Contextual definition extraction
This process automatically extracts the contextual definitions of concepts for NLU-based OL.When carrying out OL for different project BIM models, the concept definitions are sometimes available and organized in a structured form, usually including standards and glossaries.For example, the OmniClass classification system (Brodt, 2016) used in North America defines properties and upper-level property classifications.However, the textual definitions of concepts are not always annotated in BIM files.Therefore, an alternative method is provided for automatically acquiring the semantic meanings of concepts from textual resources in the AEC domain.This mimics the human learning process, where knowledge engineers consult Wikipedia or other knowledge resources if they do not know what a certain concept means.

Textual corpus preprocessing
In our methodology, all types of text files related to the AEC domain, including web pages, guidelines, manuals, articles, project records, and regulatory documents, can be incorporated into a corpus for text mining.The details of data collection are introduced in Section 4.1.The collected unstructured textual documents are first cleaned and preprocessed to form a structured and functional corpus, with all images, markups, and hyperlinks removed.Next, every paragraph in a document is split into individual sections.
To ensure that the semantics of each paragraph is rich enough, a paragraph is iteratively merged with the following paragraphs in the same document until the word number of a paragraph exceeds a threshold, which is set to 8 from our fine-tuning experiment.Ultimately, all the texts are stemmed to remove morphological affixes from words.

MWE decomposition and interpretation
Definitional question answering is a task in NLP that answers definition questions like "What is X?," which is pervasive in queries submitted to search engines (Voorhees, 2001).Analogously, our goal is to search the definition sentences of property concepts concerning certain BIM object types from the domain corpus.The challenge of our task stems from the occurrence of multiword property concepts (e.g., "default end extension length").These cohesive lexemes that cross word boundaries are known as MWEs (Baldwin et al., 2003), which sometimes do not fully appear in the domain corpus as they are too specific and too long.Furthermore, the meaning of MWEs in a domain context may not be directly interpreted by aggregating the semantics of their constituent words (Le & Jeong, 2017).The existing studies in processing MWEs mainly focus on term extraction from unstructured documents (Drymonas, 2009;Thanawala & Pareek, 2018) and vector representations (Henry et al., 2018).However, no method was found specialized for extracting contextual definitions of multiword concepts.Thus, a new algorithm was proposed to deconstruct the MWE iteratively and derive the semantic meanings of constituent concepts from the corpus to resolve these issues.
The algorithm initially tokenizes and stems the input property name n (line 3).Then the stemmed name St of the property concept is processed by the function Defi-nitionExtraction to capture the semantic meaning.If the whole MWE cannot be found in the corpus, or nothing is returned, window cropping will be used to segment the MWE and capture the semantics of meaningful partitions in the property name.The window size S is initially equal to the length of St (line 4) and iteratively reduced to one (line 6−7).Each iteration moves from the left end to the ).If the definitions dt are found (line 20), these words in the window are registered as finished words (lines 22−24) so that other windows containing these words are free of definition searching (lines 12−16).The iteration is stopped when all the words are registered, and full definitions D of MWEs are obtained.Note that when the window size is 1, the words in the stop word list (e.g., "the") contain no fetching definitions because they are not semantically meaningful.Figure 4 presents an example to demonstrate how the MWE "surface area of perimeter walls" is decomposed for definition searching.At the fourth iteration, the window size is reduced to 2, and two cropped string segments (ns) "surfac area" and "perimet wall" obtain their definition sentences with respect to the object type (o) IfcSpace.These sentences are concatenated and deemed as the definition of the overall MWE.

Automatic definition extraction
This section introduces the method for automatically extracting the definitions of property name segments (ns) from the corpus, namely, the function DefinitionExtraction introduced in Section 3.2.2.The method proposed by Cui et al. ( 2007) is adapted, which leverages soft pattern matching models (SPMMs) to extract definitions from texts flexibly.SPMMs model definition pattern matching as a probabilistic process of token sequence generation.The adapted pipeline consists of the following four steps.

Training the SPMMs
The labeled definition sentences are transformed into pattern instances for training.To avoid overfitting, the words specific to search targets assessed by a Bag-of-Words statistical relevance ranking are replaced with more general tags, including Part-of-Speech and chunking tags (Jurafsky, 2000), to avoid overfitting.After selective substitution, the sentences are cropped to collect tokens surrounding the target words.The context of the resulting pattern instances consists of the search targets and three tokens on each side of the target words.
The original work (Cui et al., 2007) proposes two kinds of SPMMs: the bigram model and the profile hidden Markov model.Our approach only uses the bigram model because of the limited training data size.The bigram model is a subclass of the N-gram model that models sequential dependencies between two adjacent tokens.Herein, the bigram probability is modeled as a linear interpolation (Manning et al., 2002) between unigram and bigram so that the conditional probability of individual tokens that occurred in specific slots can be incorporated.Both are obtained through maximum likelihood estimates in the training phase: where   (  ) means token   appearing in the slot   , and (  |  ) denotes its conditional probability; || represents the frequency of token ;   (  ) −1 ( −1 ) denotes two adjacent tokens   and  −1 that occur in slots   and  −1 ; () are the total amount of unique tokens in training data, and  is a constant to smooth the probability in case of sparse data, which is set to 2 following the original model setup (Cui et al., 2007).Deviating from the original method that multiples the probabilities at each slot, our methodology averages each term to prevent the estimated bigram probabilities from being too small to be combined with other ranking variables.The formula is presented as follows: where  denotes the bigram model; L denotes the total number of slots; and  weighs the Unigram and Bigram probabilities, which are set to 0.3 from the experiment (Cui et al., 2007).

Passage retrieval
Having trained the SPMMs, the property name segment and the corresponding object entity type (e.g., wall, space, roof, etc.) can be input to obtain the contextual definitions.
The name segment is taken as the search target and all the paragraphs in the corpus that contain the target word will then be retrieved.

Relevance ranking
The retrieved paragraphs are scored for relevance ranking based on four factors: predicted probability by SPMMs, occurring positions, domain relevance, and object relevance.Like in the training phase, the retrieved passages are transformed into pattern instances and statistically matched against previously trained patterns, which yields the estimated probability ( 1 …   ).
The position of the search target that occurred in the sentence is also considered.Since the retrieved sentences do not have a fixed length, they are split into five equivalent intervals I = [0, 0.2L, 0.4L, 0.6L, 0.8L, L], where L is the length of the sentence.The relative position of the target word can be obtained depending on which interval it belongs to.The probability of the target word occurred at an interval   is modeled as follows: where   is the number of times that the target word appeared at interval   in the training data.The denominator sums the numbers of every interval.Due to the polysemy of language (Ravin & Leacock, 2000), a target word might have multiple meanings subjected to various domains.Even though the corpus is domain-specific, the noisy content from irrelevant fields can still be retrieved.Therefore, domain relevance   is used to prioritize the definitions that originate from the AEC domain.  is obtained by calculating the cosine similarity between vector embeddings of the retrieved paragraph and domain topic words, such as "construction" and "building." Besides, even in a closed domain, the semantic meaning of property concepts can be inconsistent and subject to the type of corresponding object entities.For example, in the IFC 4 schema, the definitions of property "GrossVolume" for IfcWall and IfcSpace are different (buildingSmart International Ltd., 2019).Hence, the object relevance   is introduced to measure the relatedness between the retrieved paragraphs and the object entity type.It is defined as the proportion of the occurrence times (  ) of the surface strings of object type in the paragraph against the total length (  ) of a paragraph: The overall score of a retrieved passage is a linear combination of the above four variables and is used to rank the relevance of retrieved paragraphs.The formula of the scoring function is shown below: where  (0.1),  (0.55), (0.07), and  (0.28) are the weights of each variable.The optimal settings are found by the evolutionary optimization algorithm (Back, 1996).The objective function is defined as the sum of the ranks of ground truth definitions in the sorted list of paragraphs in the training data.Once all the retrieved paragraphs are scored, the top N-ranked paragraphs are chosen as the definition descriptions of the input name segment.If no definitions were found for the current ns, the program switches to processing the next name segment until the iteration stops.Figure 4 shows how the adapted SPMM model extracts the definition for the name segment "perimeter wall."The retrieved passages are substituted with syntactic tags and cropped to form pattern instances, which are then inputted into the trained bigram model with sequential probabilities ( 1 …   ) returned.Finally, the relevance scores of the retrieved paragraphs are calculated and the top-ranked ones are adopted.

Overview
Having obtained the contextual definitions of property concepts, the containment classification is first conducted to ensure the consistency of the ontology when new concepts are added.If the incoming concept is found to coincide with a part of the ontology, it is aligned with the most relevant SP.The overall pipeline is presented in Figure 5.
For each PP along with its definition, the SPs of the associated object are first retrieved.For feature extraction, they are then processed by a name-based SS measurement model and a definition-based SS measurement model.These features are inputted into ensemble learning (EL) models for inference.In our method, there are two alternative approaches to exploiting EL models for containment classification and PP-SP alignment: 1.The two-stage approach: An EL classifier and an EL regressor are trained separately for the two tasks.The EL classifier initially identifies whether a PP overlaps with the ontology.If yes, the PP is matched to the most relevant SP based on the EL regressor, which predicts the overall SS scores between PP and SP.The resulting SS scores are ranked, and the highest SP is selected as the matched concept.2. The one-stage approach: One EL regressor that returns SS scores between PP and SP is employed for both tasks.
If the highest SS score over all relevant SPs exceeds a threshold for the target PP, the PP is judged to be contained in the ontology.The top-ranked SP is then adopted as the matched concept.
Before PP-SP alignment in both approaches, conflict recognition is carried out to exclude the contradictory terms based on the pre-encoded logic rules.This results from the fact that although all SS measurement models can predict the semantic relatedness of names and definitions, they cannot distinguish disjoint concepts.For example, the regressor will assign a high SS score to the PP "Measured net area" and SP "NetPerimeter" because both describe the dimension information of objects.However, they are different (area and perimeter are disjoint) but are likely to mislead the model into yielding high ranks.Therefore, the disjoint terms should be identified in property names, and any conflicting SPs should be free of matching.
The specific conflict recognition method is presented in Figure 6.First, a large number of pairs of disjoint concepts are collected from the ifcOWL ontology (Pauwels & Terkaj, 2016).In OWL (Bechhofer, 2009), two or more classes are disjoint if no instance belongs to both classes, regardless of how the classes are interpreted (W3C OWL Working Group, 2012).Since IFC entities comprise prefixes and mostly contain multiple words, manual selection and simplification of disjoint classes are conducted.For instance, the pair < area, length > is acquired from the triple "ifc:IfcQuantityArea owl:disjointWith ifc:IfcQuantityLength" in the ifcOWL ontology.Consequently, 455 pairs of disjoint concepts are collected for conflict resolution.The tokens in the compared PP and SP are inspected automatically and pairwise: If disjoint terms are found, the SP is contradicted with PP and is thus omitted from matching.
The details of the EL models are illustrated as follows: Sections 3.3.2and 3.3.3introduce the name-based and definition-based SS measurement models, respectively, which work as Level-0 models for EL.Section 3.3.4demonstrates the Level-1 models for EL.

Name-based SS comparison
The first Level-0 model measures the NS between the target PP and the SPs using a word embedding model.Word embedding is a technique that maps words to vector representations (Zhang et al., 2020).This study adopts global vector (GloVe; Pennington et al., 2014), one of the stateof-the-art static word embeddings, to encode the meaning of the property name by observing the ratios of word co-occurrence probabilities in the corpus.
The GloVe embedding model is trained based on the domain corpus obtained in Section 3.2.1.Then, given a PP-SP pair, the names of PP and SP are tokenized, and all stop words are removed.A pairwise cosine similarity calculation is operated between tokens of the PP name and SP name.The overall NS score is derived by averaging all the optimal cosine similarities for each token of PP with a penalty term added, shown as follows: where argmax represents an operation to find the maximum value for a target function; GE denotes GloVe embedding function;  , denotes the ith token in the kth token list (k = 1 for PP and k = 2 for SP).N and M are the lengths of the PP and SP token lists.The second term penalized the difference in length between concept names. is a weight factor, which is set to 0.9 because the cosine distance is more influential.Ultimately, the model ranks NSs of PP with all candidate SPs, and the highest one is selected.

Definition-based SS measurement
The second Level-0 model measures the similarity between concept definitions of PPs and SPs based on sentence embeddings that map full sentences into high-dimensional vectors (Conneau et al., 2017).Sentence-Bidirectional Encoder Representations from Transformers (SBERT), a modification of the pre-trained BERT network, is used in this study as it can generate semantically meaningful sentence embeddings from texts (Reimers & Gurevych, 2019).As shown in Figure 7, SBERT conducts an average pooling operation to obtain a fixed-sized sentence embedding based on the output of BERT.It then uses Siamese network structures (Schroff et al., 2015) to optimize the weights such that the resulting sentence embeddings are semantically meaningful.A Siamese network contains two identical subnetworks sharing the same weights (Bromley et al., 1993), which are used to find the similarity of the inputs by comparing their output vectors.
In this study, a SBERT model pre-trained on the Stanford Natural Language Inference (Bowman et al., 2015) and Multi-Genre Natural Language Inference (Williams et al., 2017) dataset is adopted.A new dataset containing enormous PP-SP pairs was created with manually labeled similarity scores to fine-tune the model for our specific use.The dataset creation process is explained in Appendix A. The labeled PP-SP pairs are further processed to combine different information to feed into the SBERT network.Specifically, the name and definition were concatenated as Sentence A for PP, and the name, definition, and property set were concatenated as Sentence B for SP (see Figure 7).The whole dataset was then randomly split into a training set, a validation set, and a test set in a proportion of 4:1:1 for fine-tuning and evaluating the SBERT model.Since the outputs are continuous similarity scores of sentence pairs, the mean squared error (MSE) loss is used as the regression objective function for its advantage in penalizing large prediction errors as shown below: where n is the number of samples,   the manually labeled score, and    the predicted score.After finetuning, the SBERT model can map concept definitions into 768-dimensional vectors, thereby obtaining the definition similarity (DS) score as follows:  =   *   =  ,1  ,1 +  ,2  ,2 + ⋯ +  ,768  ,768 (9) where   ,   ∈ ℝ 768 represent the vectors of PP and SP definitions, respectively; * denotes dot production.

EL models with NS and DS combination
Based on the trained GloVe model and SBERT model, this study uses stacking ensemble models to integrate results

F I G U R E 8
The training strategy for ensemble learning-based classification and regression models.GloVe, global vector; SVM, support vector machine.from two models for concept classification and matching.Stacking generalization uses a meta-learning algorithm to learn how to best combine the predictions from multiple learners (Wolpert, 1992).It first trains Level-0 models (base models) on primary training data and then takes the outputs from Level-0 models as features to train the Level-1 model (meta-model).
In our methodology, Level-0 models are name-based and definition-based SS measurement models, as introduced in Sections 3.3.2and 3.3.3,and Level-1 meta-models adopt support vector machine (SVM) models due to the tested high-performance levels in our model-selection experiment.The training strategy is shown in Figure 8. Two Level-1 models to be trained are a classifier (applicable to the two-stage approach) for containment classification and a regressor (applicable to both the one-stage and two-stage approaches) to predict SS between PP and corresponding SPs.

Development of Level-1 classifier
The classifier is trained based on a new stacking dataset.First, a set of PPs were collected and manually labeled in terms of whether they have equivalent SPs in IFC specifications.They were subsequently input into the trained GloVe and SBERT models.The top-ranked SPs from the GloVe and SBERT models are obtained.Then, the complementary DS (  ) of the selected SP by the GloVe model and the complementary NS (  ) of the selected SP by SBERT are retrieved.Consequently, these four continuous similarity values form a 4-dimensional feature space: where   and   denote the highest NS and DS, respectively.A two-class SVM classifier was adopted for Level-1 meta-learning, and the output is a binary classification indicating either true (PP concept is contained in the ontology) or false (PP concept is not contained in the ontology).A Gaussian kernel is used to build the decision boundary, and sequential minimal optimization is taken as an optimization routine.As a result, the trained Level-1 classifier is used to infer whether a fresh PP is contained in the INLE ontology in the two-stage OL approach.

Development of Level-1 regressor
The Level-1 regressor for PP-SP matching is trained based on previously labeled PP-SP pairs.The features are changed to the NS and DS scores produced by the GloVe and SBERT models: An SVM regression model with Gaussian kernel is used as the Level-1 model to output the overall SS scores (SS ∈ [0,1]) of PP-SP pairs.To fit the best hyperplane for inferring overall values, an -insensitive loss function that penalizes predictions that are farther than  from actual values is employed.The epsilon () is set to 0.1 to sustain appropriate number of support vectors.
To this end, the trained Level-1 regressor infers the SS scores between the input PP and all related SPs to align overlapped terms in both one-stage and two-stage approaches.Take two PPs, "SlantAngle" and "EndSurfaceArea" for IfcBeam, as an example (see Figure 3).The former is predicted to match an NLName_CommonProperty instance NLN_Slope_beam so that it is added as a synonym following Path 1.By contrast, "EndSurfaceArea" is identified as a new concept in the INLE ontology.Therefore, it is enriched as an object of NLName_CustomProperty following Path 3.

Automatic population of the INLE ontology
Furthermore, as the INLE ontology is continuously enriched with various name descriptions (synonyms) of concepts, the semantic retrieval of BIM models could be enhanced by adding a logic disjunction (logic-OR) to the existing queries and inference rules to include the newly added synonyms in the part where the property name is mentioned.

3.4.2
Information alignment and BIM model updating Finally, the attributes of the repetitive PPs in the learned BIM models can be aligned with the matched SPs to improve the information standardization and retrievability of BIM models.
By retrieving and updating PP instances in IFC BIM models with existing toolkits (e.g., XBIM; Lockley et al., 2017), the names of those PPs are changed to the names of the mapped SPs, and the property sets are also changed to the sets that include the mapped SPs.Additionally, the data types of SPs are retrieved from the ontology through semantic query languages (e.g., SPARQL; DuCharme, 2013) and compared with PP instances.If different, the data types and values of PPs are changed accordingly based on BIM data augmentation programs, which could significantly correct the mistaken property data values when different AEC systems exchange building data using IFC standards.For example, the value of a PP named "Embodied Carbon" is "41.1 (kgCO 2 )" in IfcText (literals).When the property is aligned to the SP "ClimateChange," the data type is switched to IfcMassMeasure, which is a subclass of REAL (real number) in the IFC schema.Meanwhile, the numeric value "41.1" can be obtained based on a regular expression.
This operation automatically replaces dispensable customized properties in project BIM models with SPs, reducing information errors and loss during IFC-based data exchange between stakeholders and engineering software.This shows the bi-directional benefits of the proposed OL method: While project BIM models have enriched the semantic web, the quality of the BIM models themselves has also improved.

Implementation details
This study implemented all the involved models and algorithms in the Python and Java programming languages.
The collected BIM models conform to the IFC 4 schema, and all information of interest was extracted via the XBIM toolkit.The PSD and QTO files of the IFC 4 schema are used as resources for SPs.The Sklearn and Pytorch frameworks were used to train the SVM and SBERT network models, respectively.The OWL data were manipulated using the Apache Jena framework (The Apache Software Foundation, 2011).A large number of web pages, articles, manuals, and specifications related to the AEC domain were crawled and transformed into a corpus of over 4 million words (accessible in Appendix B).Sketch Engine (Lexical Computing, 2003) was used to generate this domain corpus by entering several topic words (e.g., "building," "construction").
The performance evaluation was divided into three partitions.The proposed method of contextual definitional description extraction is assessed in Section 4.2; the performance of the containment

Experiment design
In the proposed OL method, automatic contextual definition extraction plays a role in offering definitional descriptions of property concepts for NLU model inference.To evaluate its performance, a total of 327 property concepts that were not within the PSD were collected from six IFC models used in real-world construction projects in Hong Kong.The ground truth definitions of those property concepts were manually collected from Design Building Wiki (Designing Buildings Ltd, 2021) and Wikipedia.
Among the 327 concepts, 216 were used to train the definition extraction model, while the remaining 111 properties were used to evaluate the performance.Since 216 pattern instances are still insufficient for training the bigram model, an additional 500 labeled definitions from Design Building Wiki (2021) were processed into pattern instances to enlarge the training dataset.

Test results
The top 1-, 3-, and 5-ranked paragraphs were adopted as the definition of the PP.The ground truth definitions of PPs are shortened as answer nuggets consisting of a few keywords.An automatically extracted definition (AED) was confirmed as correct when it covers the answer nuggets.
The results are presented in Table 2.The accuracy rates based on the aggregation of the top 1-, 3-, and 5-ranked paragraphs are 34.23%,47.7%, and 54.05%.The more paragraphs are combined, the higher the accuracy, because the integration of more texts reduces the influence of erroneous ranking.In comparison, the original SPMM (Cui et al., 2007) was also tested.Its accuracy rates based on the top 1-, 3-, and 5-ranked paragraphs are 19.01%,27.93%, and 33.33%, respectively.This demonstrates the strength of the adjusted model in searching contextual definitions of property concepts in MWEs and in association with certain object types.Despite the moderate accuracy rate, this performance is acceptable because the role of this stage is to provide complementary knowledge resources for OL.

4.3
Evaluation of the containment classification and PP-SP alignment

Experiment design
In the proposed OL method, two-stage and one-stage ELbased approaches were proposed as two alternatives to (a) classify whether incoming PPs are contained in the INLE ontology, and (b) align the repetitive PPs with the most relevant SPs for enriching either new concepts or synonyms.The difference between the two approaches is that the two-stage method employs a classifier for containment classification and a regressor for PP-SP alignment, while the one-stage method accomplishes the two tasks using one regressor coherently.Therefore, this study first evaluated how the two approaches perform in containment classification before assessing the task of PP-SP alignment in one go.All Level-1 models are trained in a bottom-up manner.Specifically, the Level-0 models are first trained, and their outputs are taken as features for Level-1 models.Level-0 models include the GloVe model and the SBERT model.The former was trained using the corpus generated in Section 4.2.1; for the latter, a newly created dataset containing around 2160 PP-SP pairs with labeled SS scores was used for fine-tuning the SBERT networks.The details of the PP-SP dataset are illustrated in Appendix A. The trained SBERT model has a Pearson correlation of 83.22% and a Spearman's rank correlation of 79.21% against the PP-SP test set.
Afterward, the Level-1 classifiers and regressors were trained based on 216 properties from the training set and evaluated based on 111 properties from the test set created in Section 4.2.1.The concepts with manually labeled definitions (MLD) and AEDs were then tested.All datasets are accessible in Appendix B.

Two-stage approach
To study the impacts of the selection of the Level-0 models, two more Level-1 classifiers with different Level-0 mod-els were tested.Specifically, each of them takes the GloVe model and SBERT model as the only Level-0 model.The meta-learning utilizes two features (SS score and complementary score) from the ranked first SP given by the Level-0 model.
The test results of different Level-1 classifiers are presented in Figure 10.The proposed Level-1 Classifier L1 CLF (SBERT+GloVe) based on MLD yielded an optimal accuracy of 81.08%, while the accuracy of the same L1 CLF model using AED was reduced by around 5%−9%.This reduction was caused by the imperfection of the automatically retrieved definitions, as reported in Section 4.2.2, implying that a standard definition description can significantly improve the accuracy.The main errors of L1 CLF (SBERT+GloVe) stem from the PPs and SPs that depict very similar topics but have distinct usages.This causes some fresh PP concepts wrongly hypothesized as synonyms.Figure 11a provides an example, where the PP "Total glaze area" is incorrectly classified as being contained in the ontology because of the SP "Glazin-gAreaFraction," which describes a ratio rather than an area.
The Level-1 classifiers with other combinations of Level-0 models (L1 CLF (GloVe) and L1 CLF (SBERT)) perform much worse than L1 CLF (SBERT+GloVe).This is because fewer input features and biased prediction from a single Level-0 model make EL models less robust.For example, the SP "occupant" has its top-ranked PP "AreaPerOccupant" predicted by GloVe model.However, their definitions vary greatly, which leads to a wrong classification by L1 CLF (GloVe).

One-stage approach
Since the one-stage approach applies the Level-1 regressor for containment classification, other regressors that predict SS between PP and SP were compared.Two Level-0 models, namely, the GloVe model and the SBERT model, were tested in practice.All the threshold values for containment classification were obtained by optimizing the accuracy value in the test set.
As shown in Figure 12, the proposed Level-1 regressor (L1 REG) based on MLD yielded the best accuracy with 79.27%, which outperformed the other regression methods.It was found that the name variance and noisy sentences in definitions would reduce the overall SS score.In some cases, although the correct concept can be mapped, the predicted SS score is too low to pass the threshold.Figure 11b demonstrates an example that has the correct SP "Slope" as the top-ranked concept.However, the overall SS score (0.59) is less than the threshold (0.63) because of the poor NS and an overly detailed definition description for SP.
F I G U R E 1 0 Performance of different Level-1 classifiers in the two-stage approach."L1" and "CLF" denote "Level 1″ and "Classifier," respectively; the content in the bracket denotes different Level-0 models."MLD" means the manually labeled definitions; "AED top N" represents the definition from top N-ranked automatically extracted paragraphs by soft pattern matching model (SPMM).Each color denotes a tested model and filling patterns suggest the different chosen data (MLD/AED).GloVe, global vector.

F I G U R E 1 1
Examples of misdiagnosed results.The solid and dotted arrows denote predicted and ground truth answers, respectively.SP, standard property.
F I G U R E 1 2 Performance of different regressors in one-stage approach."L0," "L1," and "REG" denotes "Level 0," "Level 1," and "Regressor," respectively; "MLD" means the manually labeled definitions; the content in the bracket denotes different Level-0 models."AED top N" represents the definition from top N-ranked automatically extracted paragraphs by SPMM."PN" denotes property name.Each color denotes a tested model and filling patterns suggest the different chosen data (MLD/AED/PN).GloVe, global vector.

Comparative analysis: Two-stage approach versus one-stage approach
The two-stage and one-stage approaches were compared in a controlled environment.The condition in which the standard definitions exist was first considered.In this case, the proposed two-stage approach (L1 CLF (SBERT+GloVe)-MLD) outperforms the one-stage approach (L1 REG (SBERT+GloVe)-MLD) by 1.81%.Second, if the standard definitions are not provided, the two-stage approach based on the top 3-ranked AED (L1 CLF (SBERT+GloVe)-AED Top3) outperformed all the one-stage regression methods by at least 2.78%.
In summary, the two-stage approach outperforms the one-stage approach regardless of which standard definitions are utilized.This implies that training a Level-1 classifier is a more reliable method for containment classification.However, since the one-stage approach avoids using an additional classifier, it has the advantage of requiring less effort on model development and lower computation costs for inference.

Test results of PP-SP alignment
In this section, the proposed method of PP-SP alignment using a Level-1 SVM regressor is evaluated based on the 50 PPs that are labeled as repetitive concepts in the test set.When the SP with the top-ranked SS score returned by the regressor is consistent with the ground truth SP, the alignment for a PP is correct.Note that the assessment is independent of the containment classification results by the two-stage or one-stage approaches in the previous step.
The Level-0 regressors, including the GloVe model and SBERT model, were also tested to validate the effectiveness of stacking EL models.Moreover, since lots of relevance measure methods exist, this study selects three commonly used models as baselines: 1. Word2Vec (Peters et al., 2018) that maps property names into 100-dimensional vectors.The NS is calculated by Formula (7) for concept matching.

Term frequency-inverse document frequency (TF-IDF)
that is a popular model to rank relevance of unstructured texts.The standard definitions of PPs and SPs are transformed into 20,000-dimensional sparse matrix.The cosine similarity is calculated for concept comparison and matching.3. WordNet-based text similarity measurement model (Mihalcea et al., 2006) that predicts the SS between standard definitions of property concepts for relevance ranking.
The first two baseline models are trained on top of the corpus obtained in Section 4.2.1.A conflict recognition was operated before each model predicted the SS scores.
The test results are presented in Figure 13.Consequently, the proposed level-1 regression model based on MLD (L1 REG (SBERT+GloVe)-MLD) yielded the highest accuracy of 76%, which greatly surpassed other methods.If MLDs are not used, the L1 REG with the top 3-and top 1-ranked AED gave the best results, with an accuracy of 66%, compared with the L0 SBERT (60%−64%) and the L0 GloVe (56%).This again shows that the EL-based approach could produce better results than its constituent Level-0 models.On the contrary, all baselines performed worse than the proposed method.The Word2Vec model fails to align complex concepts that entail a grasp of professional meanings (e.g., "embodied energy" vs. "total primary energy consumption").Furthermore, the poor performance of TF-IDF and WordNet-based models suggests that the proposed deep NLU-based method is a better way to process concept definitions in OL.
There are some hardships observed that interfere with the performance of PP-SP alignment.First, the definitions of some PP-SP pairs have disparate levels of details, emphases, and text lengths.This causes the cosine distance between sentence vectors to not precisely reflect the semantic relatedness between property concepts.As shown in Figure 11c, the DS between PP ("Fresh air quantity per person") and its ground truth SP ("Out-SideAirPerPerson") is 0.56.In contrast, the score for the top-ranked SP ("HeatingDesignAirflow") is 0.64 because of the more abundant descriptions and more overlapped words (e.g., "ventilation").Second, there are many SPs in the same Pset that delineate similar characteristics of objects, which makes it error-prone to align with the expected SPs.For example, given a PP "Measured net area" for IfcSpace, it is hard to sort out the most appropriate SP among several candidates, including "NetPlannedArea," "NetCeilingArea," "NetFloorArea," and "NetWallArea." Last, there are other interesting findings.First, to explore the effects of conflict recognition, an ablation study was carried out.As a result, the accurate rates of all methods decreased between 4% and 12%, thereby confirming the importance of detecting disjoint terms.Second, for containment classification and PP-SP alignment, it can be observed that the optimal number for aggregating the ranked paragraphs as AED is 3 since both the L1 CLF-AED top 3 and the L1 REG-AED top 3 outputted the best results in the two tasks without using MLD.This can be explained by the fact that the accuracy of the top 1 AED from SPMM is too low (34.23%), while the top 5 AED has much irrelevant textual information.

Computation time
The developed algorithms were tested on a computer with an Intel i7 center processing unit (2.6 GHz), 16 gigabyte RAM, and a 64-bit Windows 10 operating system.The computation time is presented in  The total inference time for one property concept ranged from 12 to 25 s, depending on the selection of data (MLD or AED) and the approach (two-stage or one-stage).If MLD is not available, it takes approximately 12 s to automatically generate AEDs for a concept.In terms of containment classification and PP-SP alignment, the proposed two-stage and one-stage approaches took 13 and 12 s, respectively.

CASE STUDY
To demonstrate the significance of this study, a realworld case study was conducted on a conservation project featuring a heritage building (CRHKH, 2021) under preservation and renovation in Hong Kong (see Figure 14a).BIM models were built to document the newly constructed facilities digitally, which conform to Hong Kong Construction Industry Council BIM Standards (CIC, 2021; see Figure 14b).Four IFC models from different disciplines, including (a) architecture and structure, (b) fire service, (c) drainage service, and (d) air conditioning, were collected for OL.
The developed two-stage OL method was used to process the heritage BIM models.Some concepts' definitions come from official standards and software manuals (Revit Autodesk, 2019), and others lacking official interpretations acquire their definitions through automatic algorithms proposed in Section 3.2.Consequently, 593 property concepts for 17 types of IFC entities (e.g., IfcPipeSegment) were automatically learned and added to the INLE ontology (see Figure 15) in around 1.5 h without human intervention.Among these, 520 were identified as new concepts, and the remaining 73 properties were categorized as synonyms of existing SPs in the ontology, with an overall accuracy of about 80%.
To compare with the traditional manual ontology modeling approach, 20 PPs were randomly selected, and the first author manually conducted concept comparison and alignment with SPs, as well as adding new PPs to the ontology.Consequently, it took an average of 8.7 min to handle one PP, so the total time for processing 593 property concepts is estimated to be 86 h.By contrast, the OL method substantially reduces ontological knowledge modeling time by at most 98%, subject to necessary error correction time.
To validate the effectiveness of the proposed OL method, this conservation project's BIM coordinator and MEP engineer were interviewed to assess the method's practicability.The resulting ontology shown in Protégé editor (Musen & Team, 2015) and the procedure to construct it (by running OL scripts) are introduced to the interviewees.Then, they were asked five questions as shown in Table 4.
Regarding Q1, after understanding the basics of the semantic web, both interviewees recognize that an ontology and a user-friendly editor can help them conduct information modeling and management of element properties, especially for "electricity and mechanical models that have frequently updated parameters in specifications."Moreover, they both emphasize that the modeled properties in the ontology should be associated with the project BIM models.For Q2, they said that building ontologies by hand for their project is difficult and time-consuming, so it is significant to use the proposed OL method to construct an ontology effortlessly.After personally searching for concepts in the ontology editor by typing keywords, the interviewees respond to Q3 by agreeing that the resulting ontology with plentiful synonyms will allow the project team to efficiently retrieve existing properties for use in their work with BIM models, as well as increased awareness of concepts due to annotated definitions.For Q4, the interviewees were asked to evaluate 114 PP concepts learned from the fire service model.One thinks that all new concepts are useful and worth modeling, and the other picks nine concepts (e.g., "tick size") that are of low quality because their meanings cannot be understood.Finally, the value of reusing ontology model in other projects is highly recognized by both interviewees.One comments that:"this avoids the repetitive modeling of the same properties," and "concept usage can be standardized." In all, the qualitative assessment results in this case study are positive, which validates the effectiveness and practical value of the proposed OL method.

Contributions
The proposed method effectively incorporates the semantics of concept names and standard definitions to determine how a new property concept can be integrated into an IFC ontology.By applying deep NLU techniques, knowledge engineers can significantly reduce the time spent comparing the meanings of external concepts against complex IFC standards, which explains why the proposed method was far more efficient than manual ontology modeling in the case study (see Section 5).In many cases, textual definitions or descriptions are not available in BIM projects.The proposed method resolves this condition by adapting an SPMM for automatic definition searching from the domain corpus.Using AED, the performance of NLU-based approaches decreased by between 5% and 9%, but they still outperformed the commonly used namebased methods (Pan et al., 2008;Zhang & El-Gohary, 2016;Zhou & El-Gohary, 2021).This implies that NLU of concept definitions might be useful in automating diverse ontological knowledge modeling tasks in the AEC domain.Also, the extended queries and inference rules with more synonyms make it possible to identify a wider range of property instances in real-world BIM models.This will benefit many engineering applications, such as automatic code compliance checking and quantity take-off that use property concepts often.The contributions of this research are acknowledged in the following three aspects.Contribution 1.To the best of our knowledge, this is the first trial to apply OL techniques to automate the tasks of BIM ontology construction and enrichment in the AEC domain.This study proposed a novel method that automatically learns PP concepts from structured BIM models to enrich an INLE ontology with new concepts and syn-onyms.This benefits knowledge engineering in the construction industry by improving its automation level.Contribution 2. A new approach was developed to automatically extract contextual definitions of property concepts in MWE and in association with specific object types from a domain corpus.Any unknown concepts can be efficiently complemented with contextual definitions.This approach will help knowledge engineers understand the concepts and make it possible to apply the NLU technique to improve the capability of OL.Contribution 3.This study contributed a novel computation method that can predict the semantic relatedness between domain concepts based on the deep NLU of concept names and definitions.This method incorporates deep learning, NLP, and logic to capture the semantics of concepts for the end tasks of concept classification and alignment.
The experiment results show that our method of integrating reading comprehension of definitions outperforms the existing methods.

Limitations
While the research achievements are promising, several limitations should also be noted.
1.The method of automatic definition extraction is still time-consuming (12 s/PP).This is attributable to the MWE decomposition and interpretation, which must examine all property name segments until all words are registered.However, many unnecessary name segments do not need to be searched.For example, in Section 3.2.2, the name segments such as "surfac area of" and "of perimet" (see Figure 4) are apparently not names of any constituent concepts, but they still undergo operations of definition extraction.Therefore, some lexico-syntactic rules should be developed to recognize and discard meaningless name segments and accelerate the search.

The overall accuracy of the definition extraction based
on the adapted SPMM is still not at a satisfactory level.This is because the bigram model still cannot adequately deal with gaps in pattern instances (Cui et al., 2007).Any missing or extra tokens in the cropped pattern instances would lead the bigram model to make inaccurate estimations of sequential probabilities.In the future, another kind of SPMM, the profile hidden Markov model, should be implemented, along with more training data, to solve this problem.Moreover, deep learning-based approaches should be explored to improve performance.3. GloVe word embedding was employed to learn the vector representations of every single word in a corpus.However, concept names in MWEs cannot be encoded in GloVe embedding.Although this study measures the NS between property concepts by comparing each token in both names, this method is limited because the vector representation of the MWE learned from the corpus may be very different from every word.This adversely impacts the overall performance of the Level-1 models.In the future, methods for extracting and embedding MWEs in a domain corpus will be explored to improve NS measures and the overall performance of the EL-based method.4.Although the proposed two-stage and one-stage approaches for the containment classification and PP-SP alignment are efficient, their accuracy can be improved, and human intervention for consistency checking is currently unavoidable.In both approaches, the deep NLU models (SBERT networks) are only used to measure the semantic relatedness between the definition texts of PP and SP.However, the SS score is not the only indicator and feature for EL used to determine whether a new concept is equivalent to an existing concept in ontology.More importantly, the mentioned concepts, relationships, and logic in definitions could provide key evidence showing that the two concepts of PP and SP can be either identical or contradictory.Herein, two future directions are suggested to improve the performance.First, the deep NLU of definitions can be performed in an end-to-end manner, in which the model directly outputs the prediction results of concept classification and mapping.
Second, information extraction can be operated on definition texts for concept and relationship discovery, and then predictions can be made by combining NLU and ontology-based reasoning.5.This study only considers learning property concepts from BIM models.There are many other essential BIM concepts to learn, such as object types, materials, and processes.Besides, this method does not consider the enrichment of new property sets and the grouping of new properties in an appropriate Pset.In theory, Pset is an essential upper-level concept specifying the domain boundary of property concepts.From the perspective of ontology construction, this study focuses on populating the ontology with instances and synonyms but neglects the enrichment of new classes, hierarchies, and relations to the ontological body, which is indeed far more complex than the instance population (Petasis et al., 2011).Methods for abstracting and acquiring high-level knowledge from project documents will be explored in the future.Additionally, the proposed method can only populate a modular INLE ontology.More common BIM ontologies, such as ifcOWL, will be addressed in future studies.

CONCLUSION
Because of disparate project requirements and processes, enormous numbers of PPs are stored in BIM models beyond the scope of IFC specifications.For effective data management, retrieval, and reuse with BIM, it will be beneficial to automate the ontological modeling of this PP information.Unfortunately, such an automatic knowledge acquisition method does not currently exist.Therefore, this study bridges this gap by proposing a novel, deep NLUbased OL method that automatically populates PP concepts into an IFC ontology from structured BIM models.
The custom properties and quantities beyond the predefined PSD and QTO in IFC standards are extracted from BIM models.An adapted soft pattern-matching model is then utilized to extract contextual definitions of a concept from the domain corpus.Afterward, the PPs and their definitions are compared with relevant IFC property concepts and definitions to accomplish two tasks: (a) to determine whether incoming PP concepts are covered by the knowledge body to ensure the consistency of an ontology and (b) to match the most relevant SPs to those repetitive PP concepts.Specifically, the property names and definitions are transformed into word embeddings and sentence embeddings, respectively.The resulting NSs and DSs are further processed by two EL models: an SVM classifier for containment classification and an SVM regressor for property concept alignment.Finally, the extracted concepts are populated into the INLE ontology based on three learning paths, where they are either taken as new property concepts or new synonyms.The queries, inference rules and BIM models are updated in accordance with the evolution of the ontology.The performance of the proposed method was evaluated based on 327 custom property concepts collected from six real-world BIM models.The Level-1 classifiers and regressors were tested using two types of data: MLD and AED.The results showed that the two Level-1 models with MLD yielded the highest accuracy rates of 81.08% and 76% in concept classification and alignment tasks, respectively, which outperformed all Level-0 models (SBERT, GloVe).Furthermore, even without using MLD, both Level-1 models with the top 3-ranked AED still performed best.Therefore, this study has shown that incorporating reading comprehension of concept definitions into SS measures can significantly improve concept classification and mapping results, in contrast with past methods that purely calculated the similarity of concept names.
In a case study, the proposed approach was successfully applied to learning new property concepts of heritage BIM models.A total of 593 custom properties that are applied to architectural, structural, and MEP elements were automatically enriched in the INLE ontology without human intervention.The effectiveness of the OL method and the resulting ontology was validated through an interview with project practitioners.
This research contributes a novel OL method for the construction domain, which can automatically construct and enrich BIM ontologies.In the future, how to enrich an engineering ontology with more abstract knowledge (e.g., class, hierarchy, relations, and rules) for the construction domain will be explored.

F
Overview of the proposed natural language understanding (NLU)-based OL method.AEC, architecture, engineering, and construction; BIM, building information modeling; IFC, Industry Foundation Classes; SP, standard property.instanceswere manually collected based on existing online resources (e.g., WordNet and Wikipedia;Fellbaum, 2012) and were assigned different types of name variants as data properties.For example, hasorigin and hassynonym denote the original name and synonym of an NLName instance, respectively.In INLE ontology, BIM object properties are encoded under the class NLName_Property and divided into common properties and custom properties.A common property (NLName_CommonProperty) is a property encoded in PSD of IFC standards (SPs), while a custom property (NLName_CustomProperty) is any property that is beyond the scope of PSD (PPs).The following hierarchy defines which object possesses the properties.For example, properties in the Psets for IfcBeam are allocated as instances of NLName_CommonProperty_Beam.These property instances have different data properties to represent various forms of expressions (e.g., synonyms), data types, and definitions.Note that all instances of NLName_CommonProperty have already been encoded in the INLE ontology, while instances of NLName_CustomProperty are enriched by extracting information from BIM models.Figure 3 presents an overview of INLE ontology.The instance NLN_EndSurfaceArea_Beam is an example of a custom property for IfcBeam.

F
I G U R E 3 Overview of the IFC Natural Language Expression (INLE) ontology.the prefix of "NLName_ifc" for each class are omitted.QTO, quantity set definition; PSD, property set definition.

F
I G U R E 4 Example illustration of automatic definition extraction based on multiword decomposition algorithm and soft pattern matching models; "T" and "F" in registration block means "true" and "false."right end of MWE, while the words inside the window are joined into a string segment ns (line 18) and processed by DefinitionExtraction (line 19

F
The workflow of project-specific property (PP) classification and alignment based on the trained Level-1 classifier and regressor.PSD, property set definition; SP, standard property; SS, semantic similarity.

F
Bidirectional Encoder Representations from Transformers (BERT) embedding and Sentence-BERT (SBERT) structure based on a Siamese network.

F
Three paths of model-based ontology population based on classification results.BIM, building information modeling; NL, natural language; PSD, property set definition.
The next step is to automatically populate the INLE ontology with new objects and synonyms.As shown in Figure9, there are three paths to append new concepts onto INLE ontology depending on the classification results obtained in Section 3.3.4:1.Path 1: If a PP concept is classified as being encompassed in the IFC data model, the corresponding NLName_CommonProperty instance of its matched SP is located, and the PP name is automatically filled into the slot of data property hassynonym of the matched SP as its synonym.2. Path 2: If a PP concept is classified as being beyond the scope of IFC specifications, then it is identified as being an instance of NLName_CustomProperty.How-ever, provided that other BIM models have enriched the INLE ontology, it will still not be known whether such a concept overlaps existing NLName_CustomProperty instances or not.Hence, the same concept classification and alignment approach detailed in Section 3.3 is adopted here.If the PP is classified as having matched NLName_CustomProperty instances in the ontology, then it is appended as a synonym.3. Path 3: Following the process of Path 2, provided that a PP is classified as entirely new to the INLE ontology, a new object is added as an instance of NLName_CustomProperty along with a hasorigin data property value that denotes the concept's original name.

F
Performance of different models for PP-SP matching."L0," "L1," and "REG" denotes "Level 0," "Level 1," and "Regressor," respectively; "MLD" means the manually labeled definitions; "AED top N" represents the definition from top N-ranked automatically extracted paragraphs by SPMM."PN" denotes property name.Each color denotes a tested model and filling patterns suggest the different chosen data (MLD/AED/PN).Gray bars represent baseline models.GloVe, global vector.TA B L E 3 Computation time.Note: "s" and "min" denote second and minute, respectively.

F
Case study of a heritage building: (a) realistic photo taken on the renovation site and (b) IFC model visualized in the XBIM viewer.

F
I G U R E 1 5The expanded ontology through OL from the heritage Building Information Modeling (BIM) models of different disciplines.
List of abbreviations.
TA B L E 1
TA B L E 2

Table 3 .
The approach setup (model training) only took approximately 25 min.The most time-consuming step was fine-tuning the SBERT networks over thousands of PP-SP pairs, which took about 15 min.
Interview questions and answers.
TA B L E 4