Clause-Level Risk Identification in Commercial Contracts through Context-Aware Document Modeling

Abstract

Contract review requires identifying provisions that may create legal or operational exposure, yet the meaning of an individual clause often depends on definitions, exceptions, and obligations appearing elsewhere in the document. This paper investigates clause-level risk identification in commercial contracts using a context-aware document model that combines local clause representations with information retrieved from semantically related sections. The study uses 18,742 commercial agreements containing 326,000 manually reviewed clauses, covering indemnification, termination, liability, confidentiality, payment, data protection, and intellectual property provisions. Among these clauses, 41,380 are labeled as containing at least one material contractual risk. The proposed method achieves a macro-F1 of 86.4%, compared with 78.9% for an independent clause classifier and 82.1% for a long-context document baseline. The largest improvement is observed for clauses involving exceptions and cross-references, where recall increases from 68.7% to 81.5%. In a separate evaluation of 1,200 previously unseen contracts, the model reduces the median number of clauses requiring manual inspection by 37.6% while retaining 92.3% of identified high-risk provisions. Error analysis shows that ambiguous governing-law language and highly customized liability structures remain the primary sources of false predictions. These results demonstrate that explicitly modeling relationships between contract sections can improve automated risk screening beyond isolated clause classification.