Corporate internal control evaluation based on contrastive learning neural networks from the perspective of multi-source data

Traditional approaches to assessing corporate internal control effectiveness often depend on isolated data sources or simplistic fusion methods, resulting in limited accuracy and generalisability.To address this, this paper introduces a multi-view contrastive learning network framework that integrates financial statements, managerial narrative disclosures, and market sentiment data as complementary views of a firm's underlying risk profile.The model first learns aligned, robust representations across modalities through self-supervised intra-view and cross-view contrastive pre-training.It then fine-tunes on labelled data with an attention-based fusion mechanism for internal control weakness classification.Experiments show that the proposed framework achieves an area under the curve of 0.842 and a precision of 0.502, exceeding the best baseline by 2.9 percentage points in area under the curve and 10.3% in precision.These results demonstrate that the multi-view contrastive learning network framework significantly enhances the performance, robustness, and interpretability of automated internal control evaluation.