This paper addresses the challenge of accurately identifying cost factors and their interrelations for effective cost management in urban rail transit projects. Traditional manual methods are inefficient, subjective, and prone to error. Furthermore, existing Natural Language Processing (NLP) models struggle to handle the dense, intricate, and overlapping entity-relation structures characteristic of cost management texts. This study proposes a specialized automated model to jointly extract cost factor triplets and demonstrate their downstream use in a prototype system.
The study proposes a hierarchical fusion span-based model (HiFUSpan) that integrates three key layers. First, the embedding layer utilizes RoBERTa with a sliding window mechanism to segment and encode long texts. Second, the encoding layer combines a Convolutional Neural Network (CNN) with channel attention to capture local features and a Bi-directional Long Short-Term Memory (BiLSTM) network to model global dependencies. Third, the decoding layer integrates span-based entity recognition, BIO-tag (beginning-inside-outside) boundary supervision, and a Biaffine classifier to optimize triplet extraction.
Experimental results demonstrate that HiFUSpan achieves a precision of 77.48%, a recall of 77.34%, and an F1 score of 77.41%. The model exhibits robust performance in handling complex semantic environments, particularly excelling in extracting overlapping triplets, long entities and dense relational structures. Additionally, the illustrative prototype for project document processing further demonstrates the potential of the extracted cost knowledge to support knowledge querying, impact-path tracing and preliminary decision support.
This study contributes to the literature by developing a task-adaptive end-to-end joint extraction framework for urban rail transit cost management texts, which are characterised by the coexistence of long entities, dense relational chains, and overlapping triplets. The framework improves the automated extraction of cost-factor triplets from complex cost texts and highlights its considerable potential for practical application in cost management.
