Year: 2025
Type: dissertation
Abstract: Engineering and construction contracts include technical documents that guide the execution of the project, such as those related to safety, quality, planning, and budgeting.The preparation of contractual proposals is often based on Bills of Quantities (BOQs) provided by the Client, which competing companies must analyse and match with their own internal databases of tasks, resources, and costs.In practice, this process is manual, time-consuming, and prone to human error, even though the decisions made are contractually binding and have a direct impact on the economic success of the project.It is typically performed under demanding time constraints, which further increases the likelihood of oversight and mistakes.This reality exposes contracts to risks such as the execution of unforeseen additional work, cost overruns, and contractual disputes.Artificial Intelligence (AI), particularly Natural Language Processing (NLP) and Neural Networks, opens new opportunities to address these inefficiencies in the Architecture, Engineering, Construction, and Operations (AECO) industry.Despite recent advances, applying these techniques to automate the classification of tasks in BOQs still presents specific challenges, especially due to the variability of construction projects and the unstructured format of these documents.In this context, this doctoral thesis proposes an approach based on AI algorithms, focusing on transformer models and hierarchical neural networks, to automate the classification of task descriptions in BOQs.To that end, the work begins with a literature review on the state of the art regarding the use of Machine Learning (ML) techniques in cost estimation and contract analysis, identifying both limitations and opportunities.This is followed by a description of the data acquisition process from online sources using web scraping techniques, and its subsequent expansion through synthetic data generation using generative AI.Based on this data corpus, several text classification models are developed and tested, including architectures based on hierarchical Bidirectional Encoder Representations from Transformers (BERT) classifiers, using prompt engineering, and a fine-tuned "Large Language Model Meta AI" (LLaMA) for the automated classification of tasks in master-format documents.The results obtained are compared against relevant state-of-the-art benchmark models.Finally, the thesis presents the main conclusions, challenges encountered, and future research directions, while also highlighting the feasibility of implementing this technology in a practical system for the construction industry, with the aim of increasing productivity, reducing errors, and optimising administrative and predictive processes associated with cost estimation.
Author: Luís Ribeiro Jacques de Sousa
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7124823411