Multimodal BERT exploration of uncertain project resilience evolution via textual and log data joint modeling

Abstract

To address the need for dynamic project resilience evaluation under uncertainty and the difficulty of jointly modeling risk texts and execution logs, this study proposes a Multimodal-BERT joint modeling framework for analyzing project resilience evolution. The framework combines a Project-BERT text-semantic branch with a Project Log-BERT event-sequence branch and integrates the two modalities through dynamic cross-modal attention and an alignment objective. It maps text data, such as daily project reports, risk registers, change logs, issue tickets and meeting minutes, into semantic vectors; at the same time, it encodes task execution logs, API call logs, approval workflow logs and exception alert logs as event sequences. A cross-modal attention mechanism, modal alignment loss and a dual-task prediction head are also added to enable collaborative training for resilience state identification and evolutionary trend prediction. Experiments were conducted on a self-constructed multimodal project-resilience dataset containing 12 independent IT projects, 8640 cleaned text records, 599,000 cleaned log records, and 21,384 temporally aligned windows. For resilience-state recognition, Multimodal-BERT achieved an accuracy of 0.914 ± 0.006 and a macro-F1 of 0.901 ± 0.006, compared with macro-F1 values of 0.858 ± 0.007 for Project-BERT, 0.842 ± 0.009 for Log-BERT, and 0.893 ± 0.008 for Multimodal Transformer.For next-window resilience-index prediction, Multimodal-BERT achieved an MAE of 0.052 ± 0.003 and an RMSE of 0.081 ± 0.004, compared with 0.058 ± 0.004 and 0.087 ± 0.005, respectively, for Multimodal Transformer. Under the self-constructed 12-project dataset and the adopted project-level evaluation setting, the results indicate that joint modeling of management text and execution logs improves window-level state recognition and short-term resilience-index prediction; broader cross-project and cross-industry validation remains necessary. The framework provides a data-driven reference for monitoring project disturbance signals and supporting dynamic risk review, while further validation on broader cross-platform and cross-industry datasets remains necessary.