International Research Journal of Education and Technology

As the size of financial documents increases in the digital world, it is becoming necessary to adopt efficient methods of automatic processing of such documents. The existing systems of invoices processing involve either manual entry of data or rule-based OCR that may become inefficient and erroneous due to varied layouts of the documents. In this paper, we present a transformer-based technique for automatic invoice understanding and summarization involving vision-language models. Our system employs contextual information as well as layout awareness to extract structured information like the invoice number, vendor name, invoice date, and total amount. We employ state-of-the-art vision-language models including LayoutLMv3 to capture visual as well as semantic relations of the documents. In addition, we employ a transformer-based summarization model for generating useful summaries from invoices. Our experiments on benchmark datasets show that our system performs significantly better than the existing OCR-based models for invoice processing tasks.

Key Words: Invoice Processing, Transformer Models, Document Understanding, Vision-Language Models, LayoutLM, OCR, Text Summarization, Deep Learning.