Deep Fusion Matcher: A Framework for Complex Sentence Similarity with Applications in Account Payable

This paper introduces Deep Fusion Matcher, a framework for complex sentence similarity detection when dealing with fragmented and non-standard textual data. We evaluate three complementary neural architectures: dual-encoder Siamese networks, attention-based Cross-Encoders, and domain-specific fine-tuned Embedding Models. Our comprehensive testing on invoice-to-purchase order matching reveals a clear performance hierarchy, with Embedding Models outperforming Cross-Encoders, which in turn surpass Siamese Networks. We contribute a rigorous validation methodology featuring four distinct matching strategies (threshold-based, highest-score, greedy, and Hungarian algorithm optimization) alongside comprehensive evaluation metrics including F1-score, Precision@K, and Mean Reciprocal Rank. The framework incorporates drift detection mechanisms to ensure ongoing reliability in dynamic business environments. Our experimental results demonstrate that fine-tuned Embedding Models achieve 90% accuracy in matching invoice lines to purchase orders (PO), significantly improving financial reconciliation processes. Beyond financial applications, we suggest potential extensions to legal document clause matching and customer support categorization. Deep Fusion Matcher addresses the challenge of fragmented text in business documents while providing a scalable solution applicable across various industries.