In the rapidly evolving landscape of the construction industry, Artificial Intelligence (AI) is emerging and can be utilized as a pivotal force to drive significant change and innovation. As construction projects become increasingly complex, the integration of AI technologies can significantly enhance efficiency, accuracy, and decision-making throughout the project lifecycle.
AI's role in construction can extend beyond automation; there is a potential to encompass sophisticated tools and systems that revolutionize various features of the industry (e.g. contract management, design optimization, and project progress and performance)
Building on the transformative potential of AI in the construction industry, recent studies have begun to explore its applications in specific areas such as contract analysis. One notable example is a study made by a team of experts and graduate students at Pohang University of Science and Technology in Korea (by Su Jin Choi and others) on text mining in construction contracts, which investigates how AI-driven text analysis can enhance the accuracy and efficiency of contract review and management.
The Study – Definitions and Methodology:
A machine learning-based integrated decision-support system was developed to provide EPC contractors with the most efficient tool that can be utilized for risk analysis of EPC contracts. This system, called the Engineering Machine learning Automation Platform (EMAP), consists of three modules: (1) Instruction to Bids (ITB) Analysis, (2) Engineering Design, and (3) Predictive Maintenance. However, the study only focuses on the ITB analysis, specifically the conditions of contract issued to bidders. The researchers have developed 2 modules in their study, (1) the Critical Risk Check (CRC) and (2) Term Frequency Analysis (TFA) under the ITB Analysis module, and described the detailed processes of how the named-entity recognition (NER) and phrase matcher techniques were implemented and tested.
(1) The CRC module identifies and standardizes rules while automatically detecting events that align with these rules. It is particularly effective for determining the presence of risk clauses throughout the entire ITB document. On the other hand, (2) the TFA module analyzes the frequency of risk clauses and similar patterns, uncovering risks that may not be detected by the rule-based CRC module. The data is presented through visual and tabulated outputs, including charts and images.
Below is a summary of the algorithm flow of the CRC module versus that of the TFA model (Figures 1 and 2 respectively):
Figure 1: Algorithm flow of the CRC module
Figure 2: Algorithm flow of the TFA module
In the study, the modules were tested to confirm how reliable those are in actual EPC projects. The methodology used was the Proof of Concept (PoC), where the results of the risk detection accuracy and time efficiency executed by the EMAP (CRC and TFA models) were compared to the results of the risk analysis done by 2 engineers (one with 2 years of experience and the other with 5 years of experience) with reference to the results of industry experts, subject matter experts (SMEs) on EPC projects with aggregate experience of more than 50 years . The modules were tested for single keywords associated with a specific risk (single rule) and intersection of several single rules (multiple rule - keywords associated with several single rule).
The Study – Results (PoC):
1- CRC Module:
- Single Rule: Results have shown that the engineers were not able to detect all the risks (120 sentences out of the 225) where all the detected risks were validated to be true while the module extracted 230 sentences where 5 of them were false positives. Although the engineer risk extraction accuracy (100%) is relatively higher than the machine learning- based CRC automatic extraction module accuracy (89%), the engineer missed detecting 46.6% of the risks while the module did not miss any. This demonstrates the module's superior performance compared to the engineers.
- Multi-Rule: The engineer extracted 28 sentences with EPC risk, whereas the module detected 36 sentences. The module accuracy of the multi-rule is 94% which is less than the accuracy of the engineer (100%). As a result of SMEs’ verification, 34 sentences of those detected by the module were verified as risks. The engineer’s detection accuracy drops by detecting only 28 out of 34 LD risk sentences (82.3% of the risks). This indicates that the module performed better in terms of not missing any risks, even though some detected risks were false positives.
2- TFA Module:
- When comparing the results of the TFA risk module, the machine learning based automatic TFA risk extraction model can significantly reduce the risk extraction rate, and the time required for EPC risk analysis (the module being almost 24 times faster than the engineers). Additionally, the engineer was able to extract 82 occurrences for the Damages label and 4 for the fit for purposes (FFP) label, where all the identified occurrences were validated by the SMEs as true. However, in the TFA module, 98 and 6 occurrences were tagged for the Damages and the FFP respectively, and 8 and 1 keywords were determined as false positives for the Damages and the FFP respectively. The module validation results of each target label are 92% and 83% for the Damages and the FFP respectively in comparison to a 100% result validation of the results by the engineer. Notwithstanding the higher accuracy of the results by the engineer, the engineer’s performance was lower as it missed detecting 8.8% of the risks associated with the Damages and 20% of those associated with the FFP.
In conclusion, the model extraction results show higher performance than the results detected by PoC subjects (engineers), and the time required for analysis is also significantly less.
The Study - Limitations:
To start with, the module has shown a high rate of false positives (incorrectly identifying harmless sentences as risky). This suggests that the module sometimes misclassified sentences. By analyzing multiple rules at once, they are extracted with a higher percentage of plain text or meaningless sentences.
Moreover, evaluating the recognition rate for automatic extraction of the TFA module, there are cases in which similar phrases were incorrectly tagged in the case of a label with insufficient learning. To reduce these errors, learning and training through big data are necessary, which is a common problem in the rule-based information extraction model. To overcome this, there is a need for continuous updates.
In addition to the above, the module was developed exclusively for the English language. As such, further processes are necessary to collect contracts in the other languages, establishing the keywords and the rules in the database in the corresponding language, although the AI and ML technique would still work.
Furthermore, according to the validation conducted in this study, the ITB document analysis is prone to deviations depending on the capabilities of each manager or engineer.
Additionally, the outcomes produced by the module can fluctuate depending on the specific data provided by the database. This variability in input data can potentially impact the consistency and dependability of the module's results, limiting its overall reliability.
It is worth noting that the modules developed in this research paper are designed to identify risks rather than analyze or quantify them in depth. While they effectively alert users to the presence of risks, their functionality does not extend to evaluating the severity or impact of these risks. As a result, decision-makers may not have the comprehensive information needed to assess the full implications of the risks or to prioritize actions based on risk levels. To enhance their utility, future iterations of these modules should incorporate capabilities for analyzing and quantifying risks, providing a more nuanced understanding that supports better-informed decisions and more effective risk management strategies.
In light of the above, since the applicability of AI in the construction industry is very primitive and limited in its ability to align to the complexities of the construction projects nowadays, it remains essential to develop its capabilities, improve its accuracy and integrate more sophisticated analytics. Continued research and innovation will be crucial in evolving AI from its current foundational stage to a more advanced and effective tool for the construction industry.
AI is no longer a futuristic concept but a present-day reality that could transform the construction industry. As AI technology continues to evolve, its impact on construction will only grow, promising a smarter and more sustainable future for the industry. However, it remains essential to further address the limitations of the AI tools and machine learning to ensure efficiency and applicability in the industry practices, including but not limited to planning and scheduling, risk assessment, HSE management, quality control and assurance, sustainability and environmental practices, labor and workforce allocation.
Ref: Su Jin Choi, So Won Choi, Jong Hyun Kim and Eul-Bum Lee, 2021, AI and Text-Mining Applications for Analyzing Contractor’s Risk in Invitation to Bid (ITB) and Contracts for Engineering Procurement and Construction (EPC) Projects, Energies 2021, 14, 4632