<![CDATA[Introduction/Main Objectives: This study examines the integration of multimodal data fusion combining natural language processing (NLP) and financial ratio analysis to test the Obfuscation Hypothesis and detect financial statement fraud in Chinese listed companies. The topic is compelling, as information asymmetry between management and investors is exploited through linguistic manipulation. The study also seeks to improve the reliability of AI-assisted fraud detection by integrating textual narratives with quantitative financial indicators. Background Problems: Financial statement fraud becomes difficult to detect when accounting figures appear normal while accompanying narratives obscure underlying anomalies. This study therefore examines whether the discrepancy between optimistic narratives and deteriorating financial performance can serve as a useful indicator of fraud risk. Novelty: This paper introduces "Financial-Linguistic Dissonance" as a novel theoretical construct, empirically demonstrating the threshold at which rhetorical obfuscation fails to mask financial decay, unlike prior studies analyzing textual and numerical data separately. Research Methods: Employing FinBERT for textual analysis and deep learning for multimodal fusion, this study analyzes annual reports from emerging market companies, combining Fog Index, tone metrics, and profitability ratios. Finding/Results: Linguistic features account for nearly 80% of the predictive power, with Financial-Linguistic Dissonance accounting for 27.13%, demonstrating that narrative manipulation provides stronger fraud signals than financial ratios. Conclusion: The multimodal framework achieves 94.7% accuracy, suggesting its potential as an AI-assisted tool for fraud-risk screening. Corporate language can provide measurable signals as an instrument of potential financial statement fraud, with practical implications for regulators, auditors, investors, and corporate boards to strengthen fraud-risk assessment.]]>
