Abstract

Requests for Information (RFIs) are essential tools for facilitating communication among stakeholders in construction project management. They serve as formal inquiries to clarify ambiguities, resolve uncertainties, and enhance coordination between project teams, ultimately supporting effective project execution. RFIs play a critical role in maintaining workflow efficiency, reducing misinterpretations, and addressing unforeseen challenges that arise during the construction process. However, despite their importance, RFIs can sometimes highlight underlying systemic inefficiencies or anomalies that, if left unaddressed, contribute to cost overruns, schedule delays, and project quality issues.
Identifying the root causes of such anomalies within RFIs is crucial for mitigating risks and improving decision-making. Root Cause Analysis (RCA) is commonly employed to trace these issues back to their sources, allowing project managers and engineers to implement corrective measures. However, existing RCA approaches, many of which are based on event graphs, tend to rely on predefined rules or statistical correlations, which may not fully capture the dynamic and evolving nature of construction-related issues.
To address these challenges, this study proposes a novel approach that integrates human feedback-augmented reinforcement learning to enhance RCA for the event graphs of text-based RFIs. Our method leverages expert insights as a core component of the HRL loop in optimizing and improving the accuracy and reliability of causal and temporal graphs by effectively identifying and correcting errors within the graphs themselves. Importantly, our method is explicitly designed for structured RFI text data in which events have been manually pre-extracted as part of a preprocessing pipeline. Specifically, we employ hierarchical reinforcement learning (HRL) to systematically decompose the problem into multiple levels of decision-making, allowing for more structured learning and adaptation.
To validate our approach, we conduct experiments using the Causal-TimeBank dataset, a benchmark corpus annotated with explicit temporal and causal relationships. The reason for choosing this benchmark is largely due to the lack of publicly available RFIs with enriched text data and annotated causal/temporal events. Experimental results demonstrate that our method outperforms conventional RCA techniques by effectively identifying and correcting errors within causal and temporal graphs. The integration of human expertise ensures that the model remains adaptable to real-world complexities, enhancing its ability to capture nuanced relationships that might otherwise be overlooked by automated approaches. Ultimately, this work contributes to the advancement of intelligent RCA systems by combining human intuition with machine learning to create more robust, interpretable, and actionable root cause analyses in construction project management.