Contract sentence-level evaluation (Con-SEN): A sentence-level semantic engine for accurate recognition of financial contract clauses

Financial contracts under regulatory review are often characterized by excessive length, intricate clause nesting, implicit negations, and cross-sentence legal dependencies — posing significant challenges for automated compliance systems. To address these issues, we propose Contract Sentence-level Evaluation (Con-SEN), a sentence-level semantic framework designed for fine-grained clause recognition and risk-oriented interpretation in financial documents. This work introduces three core innovations. First, we construct Con-SEN-Corpus, a domain-specific dataset spanning over 30 financial sub-industries from 2010 to 2024. Each sentence is annotated along three legal dimensions — clause type, legality status, and risk level — enabling multi-dimensional supervision. Second, we develop a structure-aware encoder based on an enhanced Longformer, incorporating a sliding-window mechanism and a novel clause-anchor global attention module to capture long-range dependencies and structural hierarchies across chapters. Third, we introduce a negation polarity and regulatory keyword injection module, which improves the model’s ability to resolve adversarial logic, implicit obligations, and exemption clauses — often overlooked by general-purpose LLMs. To ensure coherent predictions across the three dimensions, we propose a multi-task consistency learning strategy that jointly optimizes clause classification, legality assessment, and risk estimation. Extensive experiments on real-world contract datasets show that Con-SEN significantly outperforms leading LLMs such as ChatGPT and Claude, achieving 18–25 percentage points higher document-level accuracy, up to 23% greater sentence coverage, and up to 6% improvements in clause-level classification. Moreover, it reduces volatility across evaluation metrics by an order of magnitude. These results position Con-SEN as a precise, regulation-aware framework for contract analysis, capable of handling the linguistic and structural complexity inherent in financial compliance tasks.