Automatic Classification of Construction Work Codes in Bill of Quantities of National Roadway Based on Text Analysis

The BOQ is a document that includes lists of brief descriptions of work items with estimated quantities and unit costs of the construction task (Seeley 1993). As shown in Table 1, each line of the WID in the BOQ requires quantities and unit cost; the cost is composed of materials, labor, expenses, and total cost. Those data are filled out by quantity surveyors and estimators in the early stage of the construction project for procurement and contract. Even after the procurement process, they can be used as base data for design changes, interim payment, and final payment (Kang and Paulson 1997). Therefore, an accurate and well-organized BOQ prepared by an experienced expert is important (Cunningham 2014) to provide adequate and precise evidence for the construction project.

In this context, the classification of the WID plays a significant role in structuring the information and maintaining the consistency of data from the design and construction stages to the maintenance stage of the project (Ekholm 1996). Cost breakdown structure is a hierarchical category that classifies all kinds of costs (Le et al. 2009). The WCC based on the cost breakdown structure is recommended to be allocated to a WID in the BOQ. The standard WCC representing the cost breakdown structure is a key index for managing cost data throughout the life cycle of a construction project. In the United Kingdom, the standard WCC was regulated in the Civil Engineering Standard Method of Measurement in 1976 and has been put down next to the detailed item in construction work. In South Korea, the Public Procurement Service published a standardized WCC to the public for standardizing BOQ in 1999 by benchmarking the UK’s systems (Yun 2010).

The hierarchy of the main class and subclasses of the Korean cost breakdown structure and examples of 12-digit WCCs are shown in Fig. 1. The first digit in the standard WCC stands for the major categories of construction work, such as architecture, civil engineering, electrical, mechanical, and cultural heritage repair. The following digits are the subdivisions of cost breakdown structure, and detailed construction works are gradually classified. Finally, one unit cost can be matched to one WID next to the standard WCC. When costs are calculated in the subdivision of construction work using the standard WCC sophistically, budget omission can be minimized, and the accuracy of cost estimation can be improved (Cerezo-Narváez et al. 2020). However, many of BOQs are often composed of their own company codes whose implied meaning is unknown, such as “TCDI21311N1” and “O108A10020,” as shown in Table 1. In this case, it is challenging to collect unit costs because there is no standard index as a basis for data collection.

Over recent years, many researchers and organizations have attempted to improve standard measurement methods for the successful management of construction projects’ costs and schedules (Cho et al. 2013; Jung and Gibson 1999). The Public Procurement Service in South Korea, which conducts public construction project contracts and price surveying, provides a system named C3R (i.e., construction cost-accounting compatible regulations) for unified BOQs. However, quantity surveyors and cost estimators do not use the standard WCC but rather submit BOQs to a contractor in a nonunified format by using their own company’s code. The situation is similar in other countries. Martínez-Rojas et al. (2016) in Spain said that the structure and linguistics of BOQs vary among practitioners without standards, making the acquisition and management of the information contained in the BOQ laborious. In Malaysia, standard measurement methods have been required to generate easy and clear BOQs (Akbar et al. 2015; Bandi et al. 2014; Rashid et al. 2006), and Akbar et al. (2015) claimed that the process should be automated and carried out in a more scientific manner. In Australia and New Zealand, the Australian and New Zealand Standard Method of Measurement (ANZSMM) was published in 2018 to provide consistent guidance for the measurement of building works. However, Senaratne and Rodrigo (2019) said that adhering to a standard method of measurement to prepare an accurate BOQ is tedious for any quantity surveyor and needs to be simplified to improve efficiency. In this context, the authors identified the necessity of an automatic assignment method that has been defined for the preparation of standardized BOQ.

In the construction management field, text classification plays a major role in document structuring, specification review, and cost calculation (Ekholm 1996). Several text classification techniques have been developed to categorize and structure the set of text information into predefined classes (Kesavaraj and Sukumaran 2013). Caldas and Soibelman (2003) proposed construction project document classification methods according to the Construction Industry Competency Standard (CICS) items. They explained the concept of the vectorization of the text for interpretation of the text features. Caldas and Soibelman (2003) also compared four well-known classification methods—naïve Bayes, $K$-nearest neighbor, Rocchio, and support vector machine algorithms—to classify documents. They applied a hierarchical classification of product documents from a construction product database according to their corresponding Construction Specifications Institute (CSI) MasterFormat division. Similarly, Al Qady and Kandil (2015) classified documents by discourse and retrieved the most relevant documents automatically from past projects. In recent years, state-of-the-art machine learning techniques have been widely used for text classification in construction. Mo et al. (2020) proposed a machine learning–based classification model that assigned and prioritized work orders for building maintenance, achieving 77% accuracy. Moon et al. (2021) developed semantic text-pairing techniques for relevant provision identification during the construction specification review process using a Doc2Vec model.

Nevertheless, experiments for automatically assigning a work classification code for a short string of descriptions in BOQs have had relatively little attention (Martínez-Rojas et al. 2016; Sequeira and Lopes 2015). Much of the research related to BOQs has focused on collaborative editing techniques and electronic document-sharing (Wang et al. 2015). However, as text-mining techniques have been actively developed, several studies have been conducted to classify construction work item description using text in BOQs. Martínez-Rojas et al. (2015) developed a mechanism for automatically extracting information from the work descriptions and storing it in a predefined hierarchical BOQ structure. The research was based on a multicriteria aggregation process, based in turn on expert knowledge and learned information. In their continuing research, they have classified the work descriptions in BOQs of residential buildings by using text classification algorithms (Martínez-Rojas et al. 2018). However, in their research, data with a lack of features were excluded from the experimental scope and classified work descriptions at the first level of the work breakdown structure. In Korea, Jeong (2020) proposed a framework for classifying work descriptions in BOQs using Jaro–Winkler distance; however, the algorithm was limited to tiling tasks only. To overcome the limitations of current BOQ text analysis studies, this research aims to assign WCCs to the lowest level of the construction work item and for this to be applied in all construction works.

This research proposes a framework to standardize BOQs by assigning a standard WCC to a detailed WID. The BOQ data used in the research were from the document prepared for the national roadway construction project in South Korea. Most of the BOQs for national roadway construction included road, bridge, and tunnel works. The research was composed of three phases, as shown in Fig. 2: (1) preprocessing, (2) similarity scoring and string matching, and (3) Evaluation.

The first phase, preprocessing of the BOQ data, proceeded on two tracks. One track was to make analysis data sets using the WIDs in unstandardized BOQs. The analysis data set comprised 44 unstandardized BOQ documents used for the procurement process for national roadway construction projects from 2016 to 2020. The project scope was confined to the value of the construction, which was more than 20 billion Korean won (approximately USD 16 million). The other track generated the standard data set using standard WIDs with standard WCCs provided by the Public Procurement Service in 2020. The document of standard WCCs contained 12,750 rows of standard WIDs. Each row represented a detailed explanation of one standard WCC.

The second phase was to score the similarity between the string from the analysis data set and the standard data set to assign the most suitable standard WCC. The similarity scoring algorithm presented in this framework scored the similarity by comparing the one detailed WID from the analysis data set to all standard WIDs from the standard data set. The most similar standard WCC with the highest similarity score was chosen and returned the final assignment result matching the WID in the analysis data set.

The final phase was the evaluation process, for checking the reliability of the framework for similarity scoring and string matching. The authors reviewed the assigned results and set the threshold of the similarity score to divide whether correctly matched or not. After the evaluation process was done, the unit cost data of the WIDs from BOQ were collected and saved for future database establishment.

A detailed explanation of each process follows in the next section. The data preprocessing, such as amplifying the string information and cleansing the unnecessary expressions, was done for better string matching performance. Next, the authors devised an automatic assignment algorithm to match the standard WCC to the WID in unstandardized BOQ. Finally, the authors stored the unit cost of each WID with a standardized BOQ in the database and discussed the different performances of the string match algorithm by the work classification.

Both analysis data sets and standard data sets were text data requiring preprocessing. A computerized approach to analyzing text is natural language processing (NLP) (Liddy 2001). In general, the first step of NLP is to feature extraction from text data. However, the original text strings of detailed WIDs from unstandardized BOQs consisted of a tiny number of words, just two or three, which was insufficient for extracting features. That is to say, one WID had a lack of information for computers to classify and assign the most proper WCC. For example, “Concrete lining” work under the “Tunnel” work division and “Concrete” under the “Road” work division were completely different types of work, but the word “Concrete” on its own could not show which division the work item was inherited from. Thus, a series of tasks were required to amplify recognizable information to improve the performance of text information extraction. The authors supplemented the text information by inheriting the words of the work name from higher levels of the WID.

The two data sets of WIDs had different forms of composition; thus, the data preprocessing was done separately. The analysis data set composed of WIDs from unstandardized BOQs was preprocessed, as shown in Fig. 3. Most of the BOQs were written in .xlsx or .xml format, and the information was written in a separate cell in the inconsistent form. Therefore, the authors had to devise an aligning method to restructure the data. As illustrated in Fig. 3, upper-level work names such as Level 1 (rows highlighted by a black straight line), Level 2 (rows highlighted by a dotted line), and Level 3 (rows highlighted by a dashed line) were attached in front of the final WID level (i.e., Level 4, highlighted in grey-blue) consecutively as a list. The process was automatically done with Python scripting using the RegEx module. If a representation meant the upper level of a work name, it was pasted to a lower level of the work name and returned in a single long string in the form of a list, as shown in Fig. 3. The words of text indicating the detailed specification of the work item such as “under 5–8 m” and “ $0.7{\mathrm{m}}^3$ excavator” were also attached to the end of the string list.

During the process of inheriting the text of the upper-level work name to the lower-level WID, unnecessary symbols (e.g., “1.”, “.1.04”, “.a”, “.a-1”) were also added. The authors coded to remove unnecessary data such as special characters or meaningless symbols, leaving only pure text. After the cleansing of unnecessary words, the tokenization process was followed (Webster and Kit 1992). Tokenization (Uysal and Gunal 2014) is the converting procedure of splitting a text into words, phrases, or other meaningful parts. More specifically, a token indicates a word, and tokenization is the act of splitting a phrase or collection of text into semantic word units (Islam and Inkpen 2008). In many cases, a space mark between the words was generally used as a delimiter to split the word units. In this research, the authors delimited the line of texts in amplified WIDs using the Open Korean Text library (OKT) in the KoNLPy package. KoNLPy is an open-source package for preprocessing Korean words, which splits the word into a stem and affix (Park and Cho 2014). After the tokenization process was completed, text information was stored in the form of a cleansed list. The analysis data set preparation for the similarity check process was thus completed.

Similar data preprocessing was performed for the standard WID data set. The standard WID data set with the standard WCC also did not have information on the upper-level work names. Hence, the text of the upper-level work name had to be added to the standard WID. The major difference in the process from the analysis data set was that the authors merged the text of upper work names according to the information of the code digits. For instance, in Fig. 4, the second digit of the standard WCC “A,” the third digit “A,” and the fourth digit “1” indicate “General work,” “Temporary work,” and “Temporary facility,” respectively. The authors replaced the digit of the code with the word intended for amplifying WID data. Also, the text of the detailed specification column in the standard WCC was attached at the end of the WID string. Tokenization was performed in the same way and stored in the form of a list.

Text similarity is determined by how close two pieces of text are both lexically and semantically (Gomma and Fahmy 2013). Gomma and Fahmy (2013) suggested three similarity measuring approaches: string, corpus, and knowledge based. The corpus-based and knowledge-based approaches determine similarity between words according to information gained from large corpora or derived from semantic networks. This is contrary to string-based approaches, which measure the similarity of string sequences or character composition. String-based approaches calculate the similarity of lexical features, which means that they do not capture semantic similarities beyond trivial levels (Hua et al. 2015; Kenter and de Rijke 2015). The text length of WID data used in this study was short and insufficient to extract semantic similarity, and thus the authors decided that string-based lexical similarity was more suitable for WID text analysis.

An edit distance has been widely used for measuring short string similarity lexically, which requires the minimum number of insertions, deletions, and substitutions to transform one string into another for similarity matching (Ristad and Yianilos 1998). The Jaro–Winkler distance, a well-known edit distance measurement, was used in this study for matching two strings and calculating their similarity (Cohen et al. 2003). First, the Jaro distance of string similarity is calculated by Eq. (1)where $|s_1|$ = length of string from the analysis data set; $|s_2|$ = length of string from the standard data set; $m$ = same number of characters; and $t$ = amount of transposition.

Jaro–Winkler similarity adds weights for prefix-matching onto the Jaro distance, as in the following [Eq. (2)]where $l$ = length of the common prefix; and $p$ = constant scaling factor. The scale of the Jaro–Winkler similarity score ranges from 0 to 1. The authors made a preliminary test of the feasibility of using this similarity measurement for the BOQ text analysis by applying it into one selected BOQ. The project had 1,370 WIDs, and, as a result, the measurement found only 50.2% of string matches with a distance of 0.7 or greater. This result implied that only 50% of WIDs were standardized and correctly matched, which is insufficient for automated lexical string-matching. The authors thus determined that the simple comparison of tokenized words included in the text would be more effective for WIDs than the edit distance calculation; the authors devised a rule-based approach to match the standard WCC to unstandardized WID by checking the inclusion of the words.

The pseudocode in Fig. 5 describes the similarity-scoring algorithm to measure the lexical similarity between two strings from each data set (i.e., one WID from the analysis data set and the other WID from the standard data set). The algorithm scored similarity based on the inclusion of words. For example, if the word “logging” exists in the WID of the analysis data set and the WID of the standard data set, the overlapping word length $7/7=1$ is added to the similarity score. In the opposite direction, if the word of the WID from the standard data set is included in the WID from the analysis data set, then the similarity score is added as equivalent to the length of the same words. This scoring system is designed to minimize the similarity score not being added when adding a score only in one direction, such as “drainage” and “drain” (i.e., added similarity $\mathrm{score}=5/8$). Each WID of the unstandardized BOQ (analysis data set) calculates the similarity score by examining the entire WID in the standard data set. Finally, the algorithm returns the optimal matching result with the highest similarity score.

After returning the WID of unstandardized BOQs with the most similar WIDs from the standard data set and the similarity score between the WIDs, the unit cost attached to the WID in the unstandardized BOQ and standard WCC stuck next to the standard WID were merged to complete the data assignment.

The assignment results were saved in .csv format. As mentioned before, the results contained the detailed WID from the analysis data set: the most similar standard WID and standard WCC, the similarity score between the two WIDs, and the unit cost of the WID. The previous work classification code assigned to the WID in the unstandardized BOQ was the same as the unlabeled data because most of the originally assigned WCCs did not follow the standard work classification code system. Cross-validation is difficult because there is no set of correct answers, and it is not a data set trained through machine learning. The arbitrary process of making correct answers and comparing them to automatically matched results was not carried out because there is room for different interpretations. The authors thus manually checked the standard WID, which was automatically assigned to detailed WIDs from unstandardized BOQs, and determined a threshold for distinguishing between two groups: accurately and not accurately assigned. The cut-off value of the morphological similarity score was set to two points based on the objective opinions of researchers at the Procurement Research Institute and related industry workers. The cut-off value was used as a threshold to evaluate the accuracy of the algorithm numerically. However, even if the similarity score were higher than the threshold value, there could be an incorrect match, or even if the similarity score were less than the threshold value, there could be correctly matched results. Therefore, the authors additionally compared the automatically matched results and discuss them in the “Results and Discussion” section. After evaluating the results manually, the authors established a unit cost database that shared the same WID to review the possibility of accurate cost estimation using the standardized WID.

The automatic assignment algorithm returned the similarity score between the detailed WID of the analysis data set and the standard WID of the standard data set. The distribution of similarity scores of the total of 65,541 WIDs is shown in Fig. 6. The average similarity score was 4.26, and the range of similarity scores varied from 0 to 21.

The left-hand side of the graph (a) in Fig. 6 shows the density distribution by the similarity score, and the right-hand side of the graph (b) shows the difference of the similarity score by construction work classification. Most of the similarity scores were concentrated on four points, and the similarity scores differed by each work classification. As shown in Table 3, the average similarity score of the “In situ concrete” classification (i.e., 5.64) was higher than that of other work classifications, in parallel with the number of WIDs. Contrary to the “In situ concrete” classification, the similarity score of the “Dams and levees” classifications was low, and the number of WID data was small. WIDs of frequently used work classifications tend to be prepared similarly to the standard WID in the BOQ more than WIDs in less-used work classifications.

Although the algorithm assigned the most similar standard WCC to the WIDs, an assigned standard WCC with a low similarity score might not represent the original intent of the WID in the unstandardized BOQ. Thus, to verify the performance of the proposed algorithm, the authors manually examined setting the threshold that distinguishes between correctly assigned and unassigned. In most instances, similarity scores of 2 or higher were correctly classified, and results with scores of less than 2 were shown as incompletely assigned. The correctly assigned results that exceeded the threshold were 91.12%; the authors set the accuracy of automatically assigning the standard WCC.

In Table 4, WIDs correctly assigned high similarity scores are organized as an example. A result with a high similarity score has very similar constituent words. WIDs assigned the same standard WCC are classified as the same work item and stored in the same database.

In Table 5, WIDs assigned low similarity scores are examined. The results show that the detailed WIDs from the unstandardized BOQ and assigned standard WIDs were not correctly matched, and the composing words were inconsistent. In several cases, there is no appropriate WCC to be assigned to represent the WID. For example, the detailed WID with the lowest similarity score (i.e., 0.25) in the first row of Table 5 represented the “provisional sum.” But this item is unnecessary for the work item description and should not be included in the WID in BOQ, and there should be no appropriate WCC that can represent the item. In addition, texts such as “material price” and “coarse aggregate” appear repeatedly in Table 5, indicating government-provided materials in Korea. The Korean government directly supplies some primary construction materials, such as cement, rebar, aggregate, ready-mixed concrete, and fume tubes, to secure a stable supply. Thus, such work items do not have proper WCCs to be allocated because they are not included in BOQ estimates. Accordingly, the authors found that results with low similarity were sometimes due to the absence of a standard WCC for WIDs. The standard code management system should complement such a standard classification code for proper cost estimation in the future.

Table 6 shows controversial results that looked correctly assigned but had a low similarity. The WID in the first row of Table 6 represented “a weak ground treatment work” and was assigned to a standard WCC representing “CCD110.” The “CCD100” class means “weak ground treatment,” but it also means “excavation and replacement.” The “CCD120” class means “weak ground treatment,” but it also means “compulsory replacement.” It seems impossible to clearly distinguish whether the WID is “CCD100” or “CCD120.” This example shows that some WIDs of the BOQ have insufficient information to select the appropriate WCCs. The work to supplement the insufficient information in the WID could be performed in future studies.

The framework allows not only automatic assignment of the standard WCC to WID in unstandardized BOQs but also the collection of the unit cost data attached in the BOQ. Unit cost collection has the potential to help comparison and verification of construction project cost. The authors selected five projects in 2019 to investigate the characteristics of the assigned data grouped in the same standard WCC. All five projects were general road construction projects, but the entire construction cost varied depending on the length and width of the road and the number of bridges and tunnels included. Table 7 shows the project information and automatic assigning results of the five BOQs. The number of correctly assigned WIDs means the assignment results had a similarity score of two or higher. The average accuracy for the five projects was 87.48%.

Table 8 shows a large difference in the number of data according to the construction classification, and there are also construction work classifications that are not used in the roadway construction BOQ. Fig. 7 is plotted to see the difference in the number of WIDs by construction work classification according to the project. The notable point is that in the case of Project 2, many WIDs were classified in classes “A: General items” and “K: Bridges.” From the results, the authors inferred that the measurement method and usage of standard work item description vary depending on the characteristics of the project or the propensity of the estimator. Also, in the current situation, the collected unit cost data using standard WCC as an index is difficult to consider as a reliable cost representing the standard WCC. This is because the difference in the amount of data for each standard WCC is large, and the number of collected unit costs is very small. Among the five projects selected in 2019, the most frequently used standard WCC was “CEE000,” accounting for 246 data out of 5,821 total WIDs. The 246 WIDs were the results after excluding the unit cost data entered in zero or negative values and also excluding the unproperly assigned result, even when the similarity score was higher than the threshold. The distribution of the unit cost is plotted in Fig. 8. The 246 unit cost data became the database of “CEE000” WCC. The standard WCC indicates “stiffener and rebar work during reinforced concrete work.” The median value of the unit cost data is 289,488 Korean won, but the unit cost ranges from almost zero to 700,000 Korean won or more. It means the cause of the difference in unit cost can be found depending on the characteristics of each project and the work (e.g., tunnel or bridge). When the unit cost data are continuously collected, they can be used to derive a more appropriate construction cost estimation.

The authors proposed the rule-based WID matching algorithm to standardize the BOQ. The concept of the algorithm involved focusing on the inclusion of words in the WID. This can help standardize detailed BOQ work by efficiently finding the most similar standard WID. However, there are some limitations.

First, the limitation in the algorithmic aspect is as follows: because only word inclusion is mainly considered, the entire length of the WID and the importance of the primary keyword cannot be considered when scoring the similarity. Second, in terms of data verification, the ultimate goal is to derive an appropriate unit price by collecting unit price data for each WID. However, there is a limit to deriving the optimal unit price for each WID using only the data collected in this paper. Because the characteristics of the project are diverse, it will be possible to extract an appropriate unit price when more data are collected in a standardized format.