Natural Language Processing with Multitask Classification for Semantic Prediction of Risk-Handling Actions in Construction Contracts

Abstract

Construction projects are capital-intensive and risk-prone, which can lead to serious claims and disputes. Thus, early identification and intervention of potential risks in contracts play significant roles in preventing conflicts in advance. However, traditional approaches are mostly limited to the simple task of predicting fragmentary information (e.g., a type of risk) from contracts. This study aims to predict comprehensive information to determine risk-handling actions by simultaneously performing three classification tasks (i.e., risk identification, risk allocation, and risk response). Specifically, the proposed multitask model is designed to integrate shared layers extracting general features for all three tasks with task-specific layers extracting relevant features of each individual task. Thus, this approach allows learning both common and specific features within a single network. For performance evaluation, experiments were performed on a data set of 2,586 contractual clauses from 10 construction projects, in which performance was compared with single-task models not only on the entire data set but also on the smaller number of data. The results revealed that the proposed model exhibited higher performance (mean weighted F1 score of 0.90 and accuracy of 0.78) than single-task models; furthermore, shared layers may better recognize hidden patterns for each classification task with the smaller data set (e.g., 0.04 higher mean F1 score and 0.09 higher accuracy for 250 samples). Thus, the proposed model can successfully implement three tasks simultaneously. When such information (e.g., risk types, responsible parties, and corresponding response strategies) is available in an early contract review, contracting parties shall determine specific risk-handling actions for proactive risk assessment and management in construction contracts.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 48 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.