Financial fraud in corporate disclosures remains a persistent and complex challenge within forensic accounting. Traditional fraud detection methods often depend on shallow lexical indicators or manual audits, which lack scalability and fail to capture the semantic depth of deceptive narratives. Addressing this gap, this research introduces a multimodal forensic accounting fraud detection system that integrates transformer-based NLP models with classical machine learning classifiers. By combining both shallow (TF-IDF, Count Vectorizer) and deep contextual representations (BERT, Longformer), the study systematically examines model–vectorizer synergies to identify optimal configurations for fraud classification. Experimental findings highlight the superiority of pairing TF-IDF with XGBoost, achieving an impressive F1-score of 97.27% and even perfect accuracy in certain validation folds. Interestingly, transformer-based embeddings yield mixed results, performing best when coupled with adaptive models such as Random Forest or Logistic Regression. These results emphasize the computational efficiency and robustness of hybrid NLP pipelines, along with their potential to enhance early fraud detection in real-world accounting contexts. The research offers a practical and scalable framework that balances interpretability, accuracy, and resource efficiency, paving the way for more effective forensic auditing technologies. However, limitations include a relatively small dataset (170 filings), a text-only focus despite the multimodal designation, and the use of transformer embeddings without task-specific fine-tuning, which may account for their lower performance.