Leveraging NLP for Semantic and Numerical Inconsistency Detection in Tax Submissions

This study presents an innovative method for identifying tax fraud through the application of natural language processing (NLP) to uncover irregularities within tax documents. Departing from conventional approaches that rely primarily on numerical analysis, the proposed framework combines domain-specific BERT embeddings with bidirectional LSTM architectures to effectively capture nuanced contextual information. A hybrid ensemble architecture is developed to process both structured data and free-text components within tax returns, facilitating the identification of semantic associations among financial entities and exposing numerical inconsistencies. The system was evaluated on a dataset comprising 15,000 tax documents, of which 8.5% were identified as fraudulent. The proposed model achieved superior performance, with an F1-score of 0.868 and an AUC of 0.931 — marking a 7.6% enhancement over leading existing models. Detection effectiveness varied by document category: individual income tax filings yielded an F1-score of 0.889, outperforming business-related filings, which scored 0.818. Further examination reveals that semantic features are particularly effective for identifying fraud in corporate tax documents, while numerical coherence indicators are more significant for personal filings. Although the approach requires higher computational resources compared to conventional techniques, its capacity to detect complex fraud schemes — especially those that disguise manipulation within textual content while maintaining plausible numeric data — offers a significant improvement to current tax fraud detection systems.